source: doc/theses/mubeen_zulfiqar_MMath/background.tex @ cc7bbe6

ADTast-experimentalenumpthread-emulationqualifiedEnum
Last change on this file since cc7bbe6 was 3a038fa, checked in by Peter A. Buhr <pabuhr@…>, 3 years ago

more updates to added text

  • Property mode set to 100644
File size: 57.4 KB
Line 
1\chapter{Background}
2
3
4A program dynamically allocates and deallocates the storage for a variable, referred to as an \newterm{object}, through calls such as @malloc@ and @free@ in C, and @new@ and @delete@ in \CC.
5Space for each allocated object comes from the dynamic-allocation zone.
6A \newterm{memory allocator} contains a complex data-structure and code that manages the layout of objects in the dynamic-allocation zone.
7The management goals are to make allocation/deallocation operations as fast as possible while densely packing objects to make efficient use of memory.
8Objects in C/\CC cannot be moved to aid the packing process, only adjacent free storage can be \newterm{coalesced} into larger free areas.
9The allocator grows or shrinks the dynamic-allocation zone to obtain storage for objects and reduce memory usage via operating-system calls, such as @mmap@ or @sbrk@ in UNIX.
10
11
12\section{Allocator Components}
13\label{s:AllocatorComponents}
14
15\VRef[Figure]{f:AllocatorComponents} shows the two important data components for a memory allocator, management and storage, collectively called the \newterm{heap}.
16The \newterm{management data} is a data structure located at a known memory address and contains all information necessary to manage the storage data.
17The management data starts with fixed-sized information in the static-data memory that flows into the dynamic-allocation memory.
18The \newterm{storage data} is composed of allocated and freed objects, and \newterm{reserved memory}.
19Allocated objects (white) are variable sized, and allocated and maintained by the program;
20\ie only the program knows the location of allocated storage, not the memory allocator.
21\begin{figure}[h]
22\centering
23\input{AllocatorComponents}
24\caption{Allocator Components (Heap)}
25\label{f:AllocatorComponents}
26\end{figure}
27Freed objects (light grey) are memory deallocated by the program, which are linked into one or more lists facilitating easy location for new allocations.
28Often the free list is chained internally so it does not consume additional storage, \ie the link fields are placed at known locations in the unused memory blocks.
29Reserved memory (dark grey) is one or more blocks of memory obtained from the operating system but not yet allocated to the program;
30if there are multiple reserved blocks, they are also chained together, usually internally.
31
32Allocated and freed objects typically have additional management data embedded within them.
33\VRef[Figure]{f:AllocatedObject} shows an allocated object with a header, trailer, and alignment padding and spacing around the object.
34The header contains information about the object, \eg size, type, etc.
35The trailer may be used to simplify an allocation implementation, \eg coalescing, and/or for security purposes to mark the end of an object.
36An object may be preceded by padding to ensure proper alignment.
37Some algorithms quantize allocation requests into distinct sizes resulting in additional spacing after objects less than the quantized value.
38When padding and spacing are necessary, neither can be used to satisfy a future allocation request while the current allocation exists.
39A free object also contains management data, \eg size, chaining, etc.
40The amount of management data for a free node defines the minimum allocation size, \eg if 16 bytes are needed for a free-list node, any allocation request less than 16 bytes must be rounded up, otherwise the free list cannot use internal chaining.
41The information in an allocated or freed object is overwritten when it transitions from allocated to freed and vice-versa by new management information and possibly data.
42
43\begin{figure}
44\centering
45\input{AllocatedObject}
46\caption{Allocated Object}
47\label{f:AllocatedObject}
48\end{figure}
49
50
51\section{Single-Threaded Memory-Allocator}
52\label{s:SingleThreadedMemoryAllocator}
53
54A single-threaded memory-allocator does not run any threads itself, but is used by a single-threaded program.
55Because the memory allocator is only executed by a single thread, concurrency issues do not exist.
56The primary issues in designing a single-threaded memory-allocator are fragmentation and locality.
57
58
59\subsection{Fragmentation}
60\label{s:Fragmentation}
61
62Fragmentation is memory requested from the operating system but not used by the program;
63hence, allocated objects are not fragmentation.
64\VRef[Figure]{f:InternalExternalFragmentation}) shows fragmentation is divided into two forms: internal or external.
65
66\begin{figure}
67\centering
68\input{IntExtFragmentation}
69\caption{Internal and External Fragmentation}
70\label{f:InternalExternalFragmentation}
71\end{figure}
72
73\newterm{Internal fragmentation} is memory space that is allocated to the program, but is not intended to be accessed by the program, such as headers, trailers, padding, and spacing around an allocated object.
74This memory is typically used by the allocator for management purposes or required by the architecture for correctness, \eg alignment.
75Internal fragmentation is problematic when management space is a significant proportion of an allocated object.
76For example, if internal fragmentation is as large as the object being managed, then the memory usage for that object is doubled.
77An allocator should strive to keep internal management information to a minimum.
78
79\newterm{External fragmentation} is all memory space reserved from the operating system but not allocated to the program~\cite{Wilson95,Lim98,Siebert00}, which includes freed objects, all external management data, and reserved memory.
80This memory is problematic in two ways: heap blowup and highly fragmented memory.
81\newterm{Heap blowup} occurs when memory freed by the program is not reused for future allocations leading to potentially unbounded external fragmentation growth~\cite{Berger00}.
82Heap blowup can occur due to allocator policies that are too restrictive in reusing freed memory and/or no coalescing of free storage.
83Memory can become \newterm{highly fragmented} after multiple allocations and deallocations of objects.
84\VRef[Figure]{f:MemoryFragmentation} shows an example of how a small block of memory fragments as objects are allocated and deallocated over time.
85Blocks of free memory become smaller and non-contiguous making them less useful in serving allocation requests.
86Memory is highly fragmented when the sizes of most free blocks are unusable.
87For example, \VRef[Figure]{f:Contiguous} and \VRef[Figure]{f:HighlyFragmented} have the same quantity of external fragmentation, but \VRef[Figure]{f:HighlyFragmented} is highly fragmented.
88If there is a request to allocate a large object, \VRef[Figure]{f:Contiguous} is more likely to be able to satisfy it with existing free memory, while \VRef[Figure]{f:HighlyFragmented} likely has to request more memory from the operating system.
89
90\begin{figure}
91\centering
92\input{MemoryFragmentation}
93\caption{Memory Fragmentation}
94\label{f:MemoryFragmentation}
95\vspace{10pt}
96\subfigure[Contiguous]{
97        \input{ContigFragmentation}
98        \label{f:Contiguous}
99} % subfigure
100        \subfigure[Highly Fragmented]{
101        \input{NonContigFragmentation}
102\label{f:HighlyFragmented}
103} % subfigure
104\caption{Fragmentation Quality}
105\label{f:FragmentationQuality}
106\end{figure}
107
108For a single-threaded memory allocator, three basic approaches for controlling fragmentation have been identified~\cite{Johnstone99}.
109The first approach is a \newterm{sequential-fit algorithm} with one list of free objects that is searched for a block large enough to fit a requested object size.
110Different search policies determine the free object selected, \eg the first free object large enough or closest to the requested size.
111Any storage larger than the request can become spacing after the object or be split into a smaller free object.
112The cost of the search depends on the shape and quality of the free list, \eg a linear versus a binary-tree free-list, a sorted versus unsorted free-list.
113
114The second approach is a \newterm{segregated} or \newterm{binning algorithm} with a set of lists for different sized freed objects.
115When an object is allocated, the requested size is rounded up to the nearest bin-size, possibly with spacing after the object.
116A binning algorithm is fast at finding free memory of the appropriate size and allocating it, since the first free object on the free list is used.
117The fewer bin-sizes, the fewer lists need to be searched and maintained;
118however, the bin sizes are less likely to closely fit the requested object size, leading to more internal fragmentation.
119The more bin-sizes, the longer the search and the less likely free objects are to be reused, leading to more external fragmentation and potentially heap blowup.
120A variation of the binning algorithm allows objects to be allocated to the requested size, but when an object is freed, it is placed on the free list of the next smallest or equal bin-size.
121For example, with bin sizes of 8 and 16 bytes, a request for 12 bytes allocates only 12 bytes, but when the object is freed, it is placed on the 8-byte bin-list.
122For subsequent requests, the bin free-lists contain objects of different sizes, ranging from one bin-size to the next (8-16 in this example), and a sequential-fit algorithm may be used to find an object large enough for the requested size on the associated bin list.
123
124The third approach is \newterm{splitting} and \newterm{coalescing algorithms}.
125When an object is allocated, if there are no free objects of the requested size, a larger free object may be split into two smaller objects to satisfy the allocation request without obtaining more memory from the operating system.
126For example, in the buddy system, a block of free memory is split into two equal chunks, one of those chunks is again split into two equal chunks, and so on until a block just large enough to fit the requested object is created.
127When an object is deallocated it is coalesced with the objects immediately before and after it in memory, if they are free, turning them into one larger object.
128Coalescing can be done eagerly at each deallocation or lazily when an allocation cannot be fulfilled.
129In all cases, coalescing increases allocation latency, hence some allocations can cause unbounded delays during coalescing.
130While coalescing does not reduce external fragmentation, the coalesced blocks improve fragmentation quality so future allocations are less likely to cause heap blowup.
131Splitting and coalescing can be used with other algorithms to avoid highly fragmented memory.
132
133
134\subsection{Locality}
135\label{s:Locality}
136
137The principle of locality recognizes that programs tend to reference a small set of data, called a working set, for a certain period of time, where a working set is composed of temporal and spatial accesses~\cite{Denning05}.
138Temporal clustering implies a group of objects are accessed repeatedly within a short time period, while spatial clustering implies a group of objects physically close together (nearby addresses) are accessed repeatedly within a short time period.
139Temporal locality commonly occurs during an iterative computation with a fix set of disjoint variables, while spatial locality commonly occurs when traversing an array.
140
141Hardware takes advantage of temporal and spatial locality through multiple levels of caching (\ie memory hierarchy).
142When an object is accessed, the memory physically located around the object is also cached with the expectation that the current and nearby objects will be referenced within a short period of time.
143For example, entire cache lines are transferred between memory and cache and entire virtual-memory pages are transferred between disk and memory.
144A program exhibiting good locality has better performance due to fewer cache misses and page faults\footnote{With the advent of large RAM memory, paging is becoming less of an issue in modern programming.}.
145
146Temporal locality is largely controlled by how a program accesses its variables~\cite{Feng05}.
147Nevertheless, a memory allocator can have some indirect influence on temporal locality and largely dictates spatial locality.
148For temporal locality, an allocator can return storage for new allocations that was just freed as these memory locations are still \emph{warm} in the memory hierarchy.
149For spatial locality, an allocator can place objects used together close together in memory, so the working set of the program fits into the fewest possible cache lines and pages.
150However, usage patterns are different for every program as is the underlying hardware memory architecture;
151hence, no general-purpose memory-allocator can provide ideal locality for every program on every computer.
152
153There are a number of ways a memory allocator can degrade locality by increasing the working set.
154For example, a memory allocator may access multiple free objects before finding one to satisfy an allocation request (\eg sequential-fit algorithm).
155If there are a (large) number of objects accessed in very different areas of memory, the allocator may perturb the program's memory hierarchy causing multiple cache or page misses~\cite{Grunwald93}.
156Another way locality can be degraded is by spatially separating related data.
157For example, in a binning allocator, objects of different sizes are allocated from different bins that may be located in different pages of memory.
158
159
160\section{Multi-Threaded Memory-Allocator}
161\label{s:MultiThreadedMemoryAllocator}
162
163A multi-threaded memory-allocator does not run any threads itself, but is used by a multi-threaded program.
164In addition to single-threaded design issues of locality and fragmentation, a multi-threaded allocator may be simultaneously accessed by multiple threads, and hence, must deal with concurrency issues such as ymutual exclusion, false sharing, and additional forms of heap blowup.
165
166
167\subsection{Mutual Exclusion}
168\label{s:MutualExclusion}
169
170\newterm{Mutual exclusion} provides sequential access to the shared management data of the heap.
171There are two performance issues for mutual exclusion.
172First is the overhead necessary to perform (at least) a hardware atomic operation every time a shared resource is accessed.
173Second is when multiple threads contend for a shared resource simultaneously, and hence, some threads must wait until the resource is released.
174Contention can be reduced in a number of ways:
175using multiple fine-grained locks versus a single lock, spreading the contention across a number of locks;
176using trylock and generating new storage if the lock is busy, yielding a classic space versus time tradeoff;
177using one of the many lock-free approaches for reducing contention on basic data-structure operations~\cite{Oyama99}.
178However, all of these approaches have degenerate cases where contention occurs.
179
180
181\subsection{False Sharing}
182\label{s:FalseSharing}
183
184False sharing is a dynamic phenomenon leading to cache thrashing.
185When two or more threads on separate CPUs simultaneously change different objects sharing a cache line, the change invalidates the other thread's associated cache, even though these threads may be uninterested in the other modified object.
186False sharing can occur in three different ways: program induced, allocator-induced active, and allocator-induced passive;
187a memory allocator can only affect the latter two.
188
189\paragraph{\newterm{Program-induced false-sharing}} occurs when one thread passes an object sharing a cache line to another thread, and both threads modify the respective objects.
190\VRef[Figure]{f:ProgramInducedFalseSharing} shows when Task$_1$ passes Object$_2$ to Task$_2$, a false-sharing situation forms when Task$_1$ modifies Object$_1$ and Task$_2$ modifies Object$_2$.
191Changes to Object$_1$ invalidate CPU$_2$'s cache line, and changes to Object$_2$ invalidate CPU$_1$'s cache line.
192
193\begin{figure}
194\centering
195\subfigure[Program-Induced False-Sharing]{
196        \input{ProgramFalseSharing}
197        \label{f:ProgramInducedFalseSharing}
198} \\
199\vspace{5pt}
200\subfigure[Allocator-Induced Active False-Sharing]{
201        \input{AllocInducedActiveFalseSharing}
202        \label{f:AllocatorInducedActiveFalseSharing}
203} \\
204\vspace{5pt}
205\subfigure[Allocator-Induced Passive False-Sharing]{
206        \input{AllocInducedPassiveFalseSharing}
207        \label{f:AllocatorInducedPassiveFalseSharing}
208} % subfigure
209\caption{False Sharing}
210\label{f:FalseSharing}
211\end{figure}
212
213\paragraph{\newterm{Allocator-induced active false-sharing}} occurs when objects are allocated within the same cache line but to different threads.
214For example, in \VRef[Figure]{f:AllocatorInducedActiveFalseSharing}, each task allocates an object and loads a cache-line of memory into its associated cache.
215Again, changes to Object$_1$ invalidate CPU$_2$'s cache line, and changes to Object$_2$ invalidate CPU$_1$'s cache line.
216
217\paragraph{\newterm{Allocator-induced passive false-sharing}} is another form of allocator-induced false-sharing caused by program-induced false-sharing.
218When an object in a program-induced false-sharing situation is deallocated, a future allocation of that object may cause passive false-sharing.
219For example, in \VRef[Figure]{f:AllocatorInducedPassiveFalseSharing}, Task$_1$ passes Object$_2$ to Task$_2$, and Task$_2$ subsequently deallocates Object$_2$.
220Allocator-induced passive false-sharing occurs when Object$_2$ is reallocated to Task$_2$ while Task$_1$ is still using Object$_1$.
221
222
223\subsection{Heap Blowup}
224\label{s:HeapBlowup}
225
226In a multi-threaded program, heap blowup can occur when memory freed by one thread is inaccessible to other threads due to the allocation strategy.
227Specific examples are presented in later sections.
228
229
230\section{Multi-Threaded Memory-Allocator Features}
231\label{s:MultiThreadedMemoryAllocatorFeatures}
232
233By analyzing a suite of existing allocators (see \VRef{s:ExistingAllocators}), the following salient features were identified:
234\begin{list}{\arabic{enumi}.}{\usecounter{enumi}\topsep=0.5ex\parsep=0pt\itemsep=0pt}
235\item multiple heaps
236\begin{list}{\alph{enumii})}{\usecounter{enumii}\topsep=0.5ex\parsep=0pt\itemsep=0pt}
237\item with or without a global heap
238\item with or without ownership
239\end{list}
240\item object containers
241\begin{list}{\alph{enumii})}{\usecounter{enumii}\topsep=0.5ex\parsep=0pt\itemsep=0pt}
242\item with or without ownership
243\item fixed or variable sized
244\item global or local free-lists
245\end{list}
246\item hybrid private/public heap
247\item allocation buffer
248\item lock-free operations
249\end{list}
250The first feature, multiple heaps, pertains to different kinds of heaps.
251The second feature, object containers, pertains to the organization of objects within the storage area.
252The remaining features apply to different parts of the allocator design or implementation.
253
254
255\section{Multiple Heaps}
256\label{s:MultipleHeaps}
257
258A single-threaded allocator has at most one thread and heap, while a multi-threaded allocator has potentially multiple threads and heaps.
259The multiple threads cause complexity, and multiple heaps are a mechanism for dealing with the complexity.
260The spectrum ranges from multiple threads using a single heap, denoted as T:1 (see \VRef[Figure]{f:SingleHeap}), to multiple threads sharing multiple heaps, denoted as T:H (see \VRef[Figure]{f:SharedHeaps}), to one thread per heap, denoted as 1:1 (see \VRef[Figure]{f:PerThreadHeap}), which is almost back to a single-threaded allocator.
261
262In the T:1 model, all threads allocate and deallocate objects from one heap.
263Memory is obtained from the freed objects or reserved memory in the heap, or from the operating system (OS);
264the heap may also return freed memory to the operating system.
265The arrows indicate the direction memory conceptually moves for each kind of operation: allocation moves memory along the path from the heap/operating-system to the user application, while deallocation moves memory along the path from the application back to the heap/operating-system.
266To safely handle concurrency, a single heap uses locking to provide mutual exclusion.
267Whether using a single lock for all heap operations or fine-grained locking for different operations, a single heap may be a significant source of contention for programs with a large amount of memory allocation.
268
269\begin{figure}
270\centering
271\subfigure[T:1]{
272%       \input{SingleHeap.pstex_t}
273        \input{SingleHeap}
274        \label{f:SingleHeap}
275} % subfigure
276\vrule
277\subfigure[T:H]{
278%       \input{MultipleHeaps.pstex_t}
279        \input{SharedHeaps}
280        \label{f:SharedHeaps}
281} % subfigure
282\vrule
283\subfigure[1:1]{
284%       \input{MultipleHeapsGlobal.pstex_t}
285        \input{PerThreadHeap}
286        \label{f:PerThreadHeap}
287} % subfigure
288\caption{Multiple Heaps, Thread:Heap Relationship}
289\end{figure}
290
291In the T:H model, each thread allocates storage from several heaps depending on certain criteria, with the goal of reducing contention by spreading allocations/deallocations across the heaps.
292The decision on when to create a new heap and which heap a thread allocates from depends on the allocator design.
293The performance goal is to reduce the ratio of heaps to threads.
294In general, locking is required, since more than one thread may concurrently access a heap during its lifetime, but contention is reduced because fewer threads access a specific heap.
295Two examples of this approach are:
296\begin{description}
297\item[heap pool:]
298Multiple heaps are managed in a pool, starting with a single or a fixed number of heaps that increase\-/decrease depending on contention\-/space issues.
299At creation, a thread is associated with a heap from the pool.
300When the thread attempts an allocation and its associated heap is locked (contention), it scans for an unlocked heap in the pool.
301If an unlocked heap is found, the thread changes its association and uses that heap.
302If all heaps are locked, the thread may create a new heap, use it, and then place the new heap into the pool;
303or the thread can block waiting for a heap to become available.
304While the heap-pool approach often minimizes the number of extant heaps, the worse case can result in more heaps than threads;
305\eg if the number of threads is large at startup with many allocations creating a large number of heaps and then the number of threads reduces.
306\item[kernel threads:]
307Each kernel thread (CPU) executing an application has its own heap.
308A thread allocates/deallocates from/to the heap of the kernel thread on which it is executing.
309Special precautions must be taken to handle or prevent the case where a thread is preempted during allocation/deallocation and restarts execution on a different kernel thread~\cite{Dice02}.
310\end{description}
311
312In the 1:1 model (thread heaps), each thread has its own heap, which eliminates contention and locking because no thread accesses another thread's heap.
313An additional benefit of thread heaps is improved locality due to better memory layout.
314As each thread only allocates from its heap, all objects for a thread are more consolidated in the storage area for that heap, better utilizing each CPUs cache and accessing fewer pages.
315In contrast, the T:H model spreads each thread's objects over a larger area in different heaps.
316Thread heaps can also eliminate allocator-induced active false-sharing, if memory is acquired so it does not overlap at crucial boundaries with memory for another thread's heap.
317For example, assume page boundaries coincide with cache line boundaries, then if a thread heap always acquires pages of memory, no two threads share a page or cache line unless pointers are passed among them.
318Hence, allocator-induced active false-sharing in \VRef[Figure]{f:AllocatorInducedActiveFalseSharing} cannot occur because the memory for thread heaps never overlaps.
319
320Threads using multiple heaps need to determine the specific heap to access for an allocation/deallocation, \ie association of thread to heap.
321A number of techniques are used to establish this association.
322The simplest approach is for each thread to have a pointer to its associated heap (or to administrative information that points to the heap), and this pointer changes if the association changes.
323For threading systems with thread-local/specific storage, the heap pointer/data is created using this mechanism;
324otherwise, the heap routines must use approaches like hashing the thread's stack-pointer or thread-id to find its associated heap.
325
326The storage management for multiple heaps is more complex than for a single heap (see \VRef[Figure]{f:AllocatorComponents}).
327\VRef[Figure]{f:MultipleHeapStorage} illustrates the general storage layout for multiple heaps.
328Allocated and free objects are labelled by the thread or heap they are associated with.
329(Links between free objects are removed for simplicity.)
330The management information in the static zone must be able to locate all heaps in the dynamic zone.
331The management information for the heaps must reside in the dynamic-allocation zone if there are a variable number.
332Each heap in the dynamic zone is composed of a list of a free objects and a pointer to its reserved memory.
333An alternative implementation is for all heaps to share one reserved memory, which requires a separate lock for the reserved storage to ensure mutual exclusion when acquiring new memory.
334Because multiple threads can allocate/free/reallocate adjacent storage, all forms of false sharing may occur.
335Other storage-management options are to use @mmap@ to set aside (large) areas of virtual memory for each heap and suballocate each heap's storage within that area.
336
337\begin{figure}
338\centering
339\input{MultipleHeapsStorage}
340\caption{Multiple-Heap Storage}
341\label{f:MultipleHeapStorage}
342\end{figure}
343
344Multiple heaps increase external fragmentation as the ratio of heaps to threads increases, which can lead to heap blowup.
345The external fragmentation experienced by a program with a single heap is now multiplied by the number of heaps, since each heap manages its own free storage and allocates its own reserved memory.
346Additionally, objects freed by one heap cannot be reused by other threads, except indirectly by returning free memory to the operating system, which can be expensive.
347(Depending on how the operating system provides dynamic storage to an application, returning storage may be difficult or impossible, \eg the contiguous @sbrk@ area in Unix.)
348In the worst case, a program in which objects are allocated from one heap but deallocated to another heap means these freed objects are never reused.
349
350Adding a \newterm{global heap} (G) attempts to reduce the cost of obtaining/returning memory among heaps (sharing) by buffering storage within the application address-space.
351Now, each heap obtains and returns storage to/from the global heap rather than the operating system.
352Storage is obtained from the global heap only when a heap allocation cannot be fulfilled, and returned to the global heap when a heap's free memory exceeds some threshold.
353Similarly, the global heap buffers this memory, obtaining and returning storage to/from the operating system as necessary.
354The global heap does not have its own thread and makes no internal allocation requests;
355instead, it uses the application thread, which called one of the multiple heaps and then the global heap, to perform operations.
356Hence, the worst-case cost of a memory operation includes all these steps.
357With respect to heap blowup, the global heap provides an indirect mechanism to move free memory among heaps, which usually has a much lower cost than interacting with the operating system to achieve the same goal and is independent of the mechanism used by the operating system to present dynamic memory to an address space.
358
359However, since any thread may indirectly perform a memory operation on the global heap, it is a shared resource that requires locking.
360A single lock can be used to protect the global heap or fine-grained locking can be used to reduce contention.
361In general, the cost is minimal since the majority of memory operations are completed without the use of the global heap.
362
363For thread heaps, when a kernel/user-thread terminates, there are two options for handling its heap.
364First is to free all objects in the heap to the global heap and destroy the thread heap.
365Second is to place the thread heap on a list of available heaps and reuse it for a new kernel/user thread in the future.
366Destroying the thread heap immediately may reduce external fragmentation sooner, since all free objects are freed to the global heap and may be reused by other threads.
367Alternatively, reusing thread heaps may improve performance if the inheriting thread makes similar allocation requests as the thread that previously held the thread heap.
368
369As multiple heaps are a key feature for a multi-threaded allocator, all further discussion assumes multiple heaps with a global heap to eliminate direct interaction with the operating system.
370
371
372\subsection{Ownership}
373\label{s:Ownership}
374
375\newterm{Ownership} defines which heap an object is returned-to on deallocation.
376If a thread returns an object to the heap it was originally allocated from, the heap has ownership of its objects.
377Alternatively, a thread can return an object to the heap it is currently allocating from, which can be any heap accessible during a thread's lifetime.
378\VRef[Figure]{f:HeapsOwnership} shows an example of multiple heaps (minus the global heap) with and without ownership.
379Again, the arrows indicate the direction memory conceptually moves for each kind of operation.
380For the 1:1 thread:heap relationship, a thread only allocates from its own heap, and without ownership, a thread only frees objects to its own heap, which means the heap is private to its owner thread and does not require any locking, called a \newterm{private heap}.
381For the T:1/T:H models with or without ownership or the 1:1 model with ownership, a thread may free objects to different heaps, which makes each heap publicly accessible to all threads, called a \newterm{public heap}.
382
383\begin{figure}
384\centering
385\subfigure[Ownership]{
386        \input{MultipleHeapsOwnership}
387} % subfigure
388\hspace{0.25in}
389\subfigure[No Ownership]{
390        \input{MultipleHeapsNoOwnership}
391} % subfigure
392\caption{Heap Ownership}
393\label{f:HeapsOwnership}
394\end{figure}
395
396\VRef[Figure]{f:MultipleHeapStorageOwnership} shows the effect of ownership on storage layout.
397(For simplicity assume the heaps all use the same size of reserves storage.)
398In contrast to \VRef[Figure]{f:MultipleHeapStorage}, each reserved area used by a heap only contains free storage for that particular heap because threads must return free objects back to the owner heap.
399Again, because multiple threads can allocate/free/reallocate adjacent storage in the same heap, all forms of false sharing may occur.
400The exception is for the 1:1 model if reserved memory does not overlap a cache-line because all allocated storage within a used area is associated with a single thread.
401In this case, there is no allocator-induced active false-sharing (see \VRef[Figure]{f:AllocatorInducedActiveFalseSharing}) because two adjacent allocated objects used by different threads cannot share a cache-line.
402As well, there is no allocator-induced passive false-sharing (see \VRef[Figure]{f:AllocatorInducedActiveFalseSharing}) because two adjacent allocated objects used by different threads cannot occur because free objects are returned to the owner heap.
403% Passive false-sharing may still occur, if delayed ownership is used (see below).
404
405\begin{figure}
406\centering
407\input{MultipleHeapsOwnershipStorage.pstex_t}
408\caption{Multiple-Heap Storage with Ownership}
409\label{f:MultipleHeapStorageOwnership}
410\end{figure}
411
412The main advantage of ownership is preventing heap blowup by returning storage for reuse by the owner heap.
413Ownership prevents the classical problem where one thread performs allocations from one heap, passes the object to another thread, and the receiving thread deallocates the object to another heap, hence draining the initial heap of storage.
414As well, allocator-induced passive false-sharing is eliminated because returning an object to its owner heap means it can never be allocated to another thread.
415For example, in \VRef[Figure]{f:AllocatorInducedPassiveFalseSharing}, the deallocation by Task$_2$ returns Object$_2$ back to Task$_1$'s heap;
416hence a subsequent allocation by Task$_2$ cannot return this storage.
417The disadvantage of ownership is deallocating to another task's heap so heaps are no longer private and require locks to provide safe concurrent access.
418
419Object ownership can be immediate or delayed, meaning objects may be returned to the owner heap immediately at deallocation or after some delay.
420A thread may delay the return by storing objects it does not own on a separate free list.
421Delaying can improve performance by batching objects for return to their owner heap and possibly reallocating these objects if storage runs out on the current heap.
422However, reallocation can result in passive false-sharing.
423For example, in \VRef[Figure]{f:AllocatorInducedPassiveFalseSharing}, Object$_2$ may be deallocated to Task$_2$'s heap initially.
424If Task$_2$ reallocates Object$_2$ before it is returned to its owner heap, then passive false-sharing may occur.
425
426
427\section{Object Containers}
428\label{s:ObjectContainers}
429
430One approach for managing objects places headers/trailers around individual objects, meaning memory adjacent to the object is reserved for object-management information, as shown in \VRef[Figure]{f:ObjectHeaders}.
431However, this approach leads to poor cache usage, since only a portion of the cache line is holding useful information from the program's perspective.
432Spatial locality is also negatively affected.
433While the header and object are together in memory, they are generally not accessed together;
434\eg the object is accessed by the program when it is allocated, while the header is accessed by the allocator when the object is free.
435This difference in usage patterns can lead to poor cache locality~\cite{Feng05}.
436Additionally, placing headers on individual objects can lead to redundant management information.
437For example, if a header stores only the object size, then all objects with the same size have identical headers.
438
439\begin{figure}
440\centering
441\subfigure[Object Headers]{
442        \input{ObjectHeaders}
443        \label{f:ObjectHeaders}
444} % subfigure
445\\
446\subfigure[Object Container]{
447        \input{Container}
448        \label{f:ObjectContainer}
449} % subfigure
450\caption{Header Placement}
451\label{f:HeaderPlacement}
452\end{figure}
453
454An alternative approach for managing objects factors common header/trailer information to a separate location in memory and organizes associated free storage into blocks called \newterm{object containers} (\newterm{superblocks} in~\cite{Berger00}), as in \VRef[Figure]{f:ObjectContainer}.
455The header for the container holds information necessary for all objects in the container;
456a trailer may also be used at the end of the container.
457Similar to the approach described for thread heaps in \VRef{s:MultipleHeaps}, if container boundaries do not overlap with memory of another container at crucial boundaries and all objects in a container are allocated to the same thread, allocator-induced active false-sharing is avoided.
458
459The difficulty with object containers lies in finding the object header/trailer given only the object address, since that is normally the only information passed to the deallocation operation.
460One way to do this is to start containers on aligned addresses in memory, then truncate the lower bits of the object address to obtain the header address (or round up and subtract the trailer size to obtain the trailer address).
461For example, if an object at address 0xFC28\,EF08 is freed and containers are aligned on 64\,KB (0x0001\,0000) addresses, then the container header is at 0xFC28\,0000.
462
463Normally, a container has homogeneous objects of fixed size, with fixed information in the header that applies to all container objects (\eg object size and ownership).
464This approach greatly reduces internal fragmentation since far fewer headers are required, and potentially increases spatial locality as a cache line or page holds more objects since the objects are closer together due to the lack of headers.
465However, although similar objects are close spatially within the same container, different sized objects are further apart in separate containers.
466Depending on the program, this may or may not improve locality.
467If the program uses several objects from a small number of containers in its working set, then locality is improved since fewer cache lines and pages are required.
468If the program uses many containers, there is poor locality, as both caching and paging increase.
469Another drawback is that external fragmentation may be increased since containers reserve space for objects that may never be allocated by the program, \ie there are often multiple containers for each size only partially full.
470However, external fragmentation can be reduced by using small containers.
471
472Containers with heterogeneous objects implies different headers describing them, which complicates the problem of locating a specific header solely by an address.
473A couple of solutions can be used to implement containers with heterogeneous objects.
474However, the problem with allowing objects of different sizes is that the number of objects, and therefore headers, in a single container is unpredictable.
475One solution allocates headers at one end of the container, while allocating objects from the other end of the container;
476when the headers meet the objects, the container is full.
477Freed objects cannot be split or coalesced since this causes the number of headers to change.
478The difficulty in this strategy remains in finding the header for a specific object;
479in general, a search is necessary to find the object's header among the container headers.
480A second solution combines the use of container headers and individual object headers.
481Each object header stores the object's heterogeneous information, such as its size, while the container header stores the homogeneous information, such as the owner when using ownership.
482This approach allows containers to hold different types of objects, but does not completely separate headers from objects.
483The benefit of the container in this case is to reduce some redundant information that is factored into the container header.
484
485In summary, object containers trade off internal fragmentation for external fragmentation by isolating common administration information to remove/reduce internal fragmentation, but at the cost of external fragmentation as some portion of a container may not be used and this portion is unusable for other kinds of allocations.
486A consequence of this tradeoff is its effect on spatial locality, which can produce positive or negative results depending on program access-patterns.
487
488
489\subsection{Container Ownership}
490\label{s:ContainerOwnership}
491
492Without ownership, objects in a container are deallocated to the heap currently associated with the thread that frees the object.
493Thus, different objects in a container may be on different heap free-lists (see \VRef[Figure]{f:ContainerNoOwnershipFreelist}).
494With ownership, all objects in a container belong to the same heap (see \VRef[Figure]{f:ContainerOwnershipFreelist}), so ownership of an object is determined by the container owner.
495If multiple threads can allocate/free/reallocate adjacent storage in the same heap, all forms of false sharing may occur.
496Only with the 1:1 model and ownership is active and passive false-sharing avoided (see \VRef{s:Ownership}).
497Passive false-sharing may still occur, if delayed ownership is used.
498
499\begin{figure}
500\centering
501\subfigure[No Ownership]{
502        \input{ContainerNoOwnershipFreelist}
503        \label{f:ContainerNoOwnershipFreelist}
504} % subfigure
505\vrule
506\subfigure[Ownership]{
507        \input{ContainerOwnershipFreelist}
508        \label{f:ContainerOwnershipFreelist}
509} % subfigure
510\caption{Free-list Structure with Container Ownership}
511\end{figure}
512
513A fragmented heap has multiple containers that may be partially or completely free.
514A completely free container can become reserved storage and be reset to allocate objects of a new size.
515When a heap reaches a threshold of free objects, it moves some free storage to the global heap for reuse to prevent heap blowup.
516Without ownership, when a heap frees objects to the global heap, individual objects must be passed, and placed on the global-heap's free-list.
517Containers cannot be freed to the global heap unless completely free because
518
519When a container changes ownership, the ownership of all objects within it change as well.
520Moving a container involves moving all objects on the heap's free-list in that container to the new owner.
521This approach can reduce contention for the global heap, since each request for objects from the global heap returns a container rather than individual objects.
522
523Additional restrictions may be applied to the movement of containers to prevent active false-sharing.
524For example, in \VRef[Figure]{f:ContainerFalseSharing1}, a container being used by Task$_1$ changes ownership, through the global heap.
525In \VRef[Figure]{f:ContainerFalseSharing2}, when Task$_2$ allocates an object from the newly acquired container it is actively false-sharing even though no objects are passed among threads.
526Note, once the object is freed by Task$_1$, no more false sharing can occur until the container changes ownership again.
527To prevent this form of false sharing, container movement may be restricted to when all objects in the container are free.
528One implementation approach that increases the freedom to return a free container to the operating system involves allocating containers using a call like @mmap@, which allows memory at an arbitrary address to be returned versus only storage at the end of the contiguous @sbrk@ area.
529
530\begin{figure}
531\centering
532\subfigure[]{
533        \input{ContainerFalseSharing1}
534        \label{f:ContainerFalseSharing1}
535} % subfigure
536\subfigure[]{
537        \input{ContainerFalseSharing2}
538        \label{f:ContainerFalseSharing2}
539} % subfigure
540\caption{Active False-Sharing using Containers}
541\label{f:ActiveFalseSharingContainers}
542\end{figure}
543
544Using containers with ownership increases external fragmentation since a new container for a requested object size must be allocated separately for each thread requesting it.
545In \VRef[Figure]{f:ExternalFragmentationContainerOwnership}, using object ownership allocates 80\% more space than without ownership.
546
547\begin{figure}
548\centering
549\subfigure[No Ownership]{
550        \input{ContainerNoOwnership}
551} % subfigure
552\\
553\subfigure[Ownership]{
554        \input{ContainerOwnership}
555} % subfigure
556\caption{External Fragmentation with Container Ownership}
557\label{f:ExternalFragmentationContainerOwnership}
558\end{figure}
559
560
561\subsection{Container Size}
562\label{s:ContainerSize}
563
564One way to control the external fragmentation caused by allocating a large container for a small number of requested objects is to vary the size of the container.
565As described earlier, container boundaries need to be aligned on addresses that are a power of two to allow easy location of the header (by truncating lower bits).
566Aligning containers in this manner also determines the size of the container.
567However, the size of the container has different implications for the allocator.
568
569The larger the container, the fewer containers are needed, and hence, the fewer headers need to be maintained in memory, improving both internal fragmentation and potentially performance.
570However, with more objects in a container, there may be more objects that are unallocated, increasing external fragmentation.
571With smaller containers, not only are there more containers, but a second new problem arises where objects are larger than the container.
572In general, large objects, \eg greater than 64\,KB, are allocated directly from the operating system and are returned immediately to the operating system to reduce long-term external fragmentation.
573If the container size is small, \eg 1\,KB, then a 1.5\,KB object is treated as a large object, which is likely to be inappropriate.
574Ideally, it is best to use smaller containers for smaller objects, and larger containers for medium objects, which leads to the issue of locating the container header.
575
576In order to find the container header when using different sized containers, a super container is used (see~\VRef[Figure]{f:SuperContainers}).
577The super container spans several containers, contains a header with information for finding each container header, and starts on an aligned address.
578Super-container headers are found using the same method used to find container headers by dropping the lower bits of an object address.
579The containers within a super container may be different sizes or all the same size.
580If the containers in the super container are different sizes, then the super-container header must be searched to determine the specific container for an object given its address.
581If all containers in the super container are the same size, \eg 16KB, then a specific container header can be found by a simple calculation.
582The free space at the end of a super container is used to allocate new containers.
583
584\begin{figure}
585\centering
586\input{SuperContainers}
587% \includegraphics{diagrams/supercontainer.eps}
588\caption{Super Containers}
589\label{f:SuperContainers}
590\end{figure}
591
592Minimal internal and external fragmentation is achieved by having as few containers as possible, each being as full as possible.
593It is also possible to achieve additional benefit by using larger containers for popular small sizes, as it reduces the number of containers with associated headers.
594However, this approach assumes it is possible for an allocator to determine in advance which sizes are popular.
595Keeping statistics on requested sizes allows the allocator to make a dynamic decision about which sizes are popular.
596For example, after receiving a number of allocation requests for a particular size, that size is considered a popular request size and larger containers are allocated for that size.
597If the decision is incorrect, larger containers than necessary are allocated that remain mostly unused.
598A programmer may be able to inform the allocator about popular object sizes, using a mechanism like @mallopt@, in order to select an appropriate container size for each object size.
599
600
601\subsection{Container Free-Lists}
602\label{s:containersfreelists}
603
604The container header allows an alternate approach for managing the heap's free-list.
605Rather than maintain a global free-list throughout the heap (see~\VRef[Figure]{f:GlobalFreeListAmongContainers}), the containers are linked through their headers and only the local free objects within a container are linked together (see~\VRef[Figure]{f:LocalFreeListWithinContainers}).
606Note, maintaining free lists within a container assumes all free objects in the container are associated with the same heap;
607thus, this approach only applies to containers with ownership.
608
609This alternate free-list approach can greatly reduce the complexity of moving all freed objects belonging to a container to another heap.
610To move a container using a global free-list, as in \VRef[Figure]{f:GlobalFreeListAmongContainers}, the free list is first searched to find all objects within the container.
611Each object is then removed from the free list and linked together to form a local free-list for the move to the new heap.
612With local free-lists in containers, as in \VRef[Figure]{f:LocalFreeListWithinContainers}, the container is simply removed from one heap's free list and placed on the new heap's free list.
613Thus, when using local free-lists, the operation of moving containers is reduced from $O(N)$ to $O(1)$.
614The cost is adding information to a header, which increases the header size, and therefore internal fragmentation.
615
616\begin{figure}
617\centering
618\subfigure[Global Free-List Among Containers]{
619        \input{FreeListAmongContainers}
620        \label{f:GlobalFreeListAmongContainers}
621} % subfigure
622\hspace{0.25in}
623\subfigure[Local Free-List Within Containers]{
624        \input{FreeListWithinContainers}
625        \label{f:LocalFreeListWithinContainers}
626} % subfigure
627\caption{Container Free-List Structure}
628\label{f:ContainerFreeListStructure}
629\end{figure}
630
631When all objects in the container are the same size, a single free-list is sufficient.
632However, when objects in the container are different size, the header needs a free list for each size class when using a binning allocation algorithm, which can be a significant increase in the container-header size.
633The alternative is to use a different allocation algorithm with a single free-list, such as a sequential-fit allocation-algorithm.
634
635
636\subsection{Hybrid Private/Public Heap}
637\label{s:HybridPrivatePublicHeap}
638
639Section~\Vref{s:Ownership} discusses advantages and disadvantages of public heaps (T:H model and with ownership) and private heaps (thread heaps with ownership).
640For thread heaps with ownership, it is possible to combine these approaches into a hybrid approach with both private and public heaps (see~\VRef[Figure]{f:HybridPrivatePublicHeap}).
641The main goal of the hybrid approach is to eliminate locking on thread-local allocation/deallocation, while providing ownership to prevent heap blowup.
642In the hybrid approach, a task first allocates from its private heap and second from its public heap if no free memory exists in the private heap.
643Similarly, a task first deallocates an object its private heap, and second to the public heap.
644Both private and public heaps can allocate/deallocate to/from the global heap if there is no free memory or excess free memory, although an implementation may choose to funnel all interaction with the global heap through one of the heaps.
645Note, deallocation from the private to the public (dashed line) is unlikely because there is no obvious advantages unless the public heap provides the only interface to the global heap.
646Finally, when a task frees an object it does not own, the object is either freed immediately to its owner's public heap or put in the freeing task's private heap for delayed ownership, which allows the freeing task to temporarily reuse an object before returning it to its owner or batch objects for an owner heap into a single return.
647
648\begin{figure}
649\centering
650\input{PrivatePublicHeaps.pstex_t}
651\caption{Hybrid Private/Public Heap for Per-thread Heaps}
652\label{f:HybridPrivatePublicHeap}
653% \vspace{10pt}
654% \input{RemoteFreeList.pstex_t}
655% \caption{Remote Free-List}
656% \label{f:RemoteFreeList}
657\end{figure}
658
659As mentioned, an implementation may have only one heap deal with the global heap, so the other heap can be simplified.
660For example, if only the private heap interacts with the global heap, the public heap can be reduced to a lock-protected free-list of objects deallocated by other threads due to ownership, called a \newterm{remote free-list}.
661To avoid heap blowup, the private heap allocates from the remote free-list when it reaches some threshold or it has no free storage.
662Since the remote free-list is occasionally cleared during an allocation, this adds to that cost.
663Clearing the remote free-list is $O(1)$ if the list can simply be added to the end of the private-heap's free-list, or $O(N)$ if some action must be performed for each freed object.
664
665If only the public heap interacts with other threads and the global heap, the private heap can handle thread-local allocations and deallocations without locking.
666In this scenario, the private heap must deallocate storage after reaching a certain threshold to the public heap (and then eventually to the global heap from the public heap) or heap blowup can occur.
667If the public heap does the major management, the private heap can be simplified to provide high-performance thread-local allocations and deallocations.
668
669The main disadvantage of each thread having both a private and public heap is the complexity of managing two heaps and their interactions in an allocator.
670Interestingly, heap implementations often focus on either a private or public heap, giving the impression a single versus a hybrid approach is being used.
671In many case, the hybrid approach is actually being used, but the simpler heap is just folded into the complex heap, even though the operations logically belong in separate heaps.
672For example, a remote free-list is actually a simple public-heap, but may be implemented as an integral component of the complex private-heap in an allocator, masking the presence of a hybrid approach.
673
674
675\section{Allocation Buffer}
676\label{s:AllocationBuffer}
677
678An allocation buffer is reserved memory (see~\VRef{s:AllocatorComponents}) not yet allocated to the program, and is used for allocating objects when the free list is empty.
679That is, rather than requesting new storage for a single object, an entire buffer is requested from which multiple objects are allocated later.
680Both any heap may use an allocation buffer, resulting in allocation from the buffer before requesting objects (containers) from the global heap or operating system, respectively.
681The allocation buffer reduces contention and the number of global/operating-system calls.
682For coalescing, a buffer is split into smaller objects by allocations, and recomposed into larger buffer areas during deallocations.
683
684Allocation buffers are useful initially when there are no freed objects in a heap because many allocations usually occur when a thread starts.
685Furthermore, to prevent heap blowup, objects should be reused before allocating a new allocation buffer.
686Thus, allocation buffers are often allocated more frequently at program/thread start, and then their use often diminishes.
687
688Using an allocation buffer with a thread heap avoids active false-sharing, since all objects in the allocation buffer are allocated to the same thread.
689For example, if all objects sharing a cache line come from the same allocation buffer, then these objects are allocated to the same thread, avoiding active false-sharing.
690Active false-sharing may still occur if objects are freed to the global heap and reused by another heap.
691
692Allocation buffers may increase external fragmentation, since some memory in the allocation buffer may never be allocated.
693A smaller allocation buffer reduces the amount of external fragmentation, but increases the number of calls to the global heap or operating system.
694The allocation buffer also slightly increases internal fragmentation, since a pointer is necessary to locate the next free object in the buffer.
695
696The unused part of a container, neither allocated or freed, is an allocation buffer.
697For example, when a container is created, rather than placing all objects within the container on the free list, the objects form an allocation buffer and are allocated from the buffer as allocation requests are made.
698This lazy method of constructing objects is beneficial in terms of paging and caching.
699For example, although an entire container, possibly spanning several pages, is allocated from the operating system, only a small part of the container is used in the working set of the allocator, reducing the number of pages and cache lines that are brought into higher levels of cache.
700
701
702\section{Lock-Free Operations}
703\label{s:LockFreeOperations}
704
705A lock-free algorithm guarantees safe concurrent-access to a data structure, so that at least one thread can make progress in the system, but an individual task has no bound to execution, and hence, may starve~\cite[pp.~745--746]{Herlihy93}.
706% A wait-free algorithm puts a finite bound on the number of steps any thread takes to complete an operation, so an individual task cannot starve
707Lock-free operations can be used in an allocator to reduce or eliminate the use of locks.
708Locks are a problem for high contention or if the thread holding the lock is preempted and other threads attempt to use that lock.
709With respect to the heap, these situations are unlikely unless all threads makes extremely high use of dynamic-memory allocation, which can be an indication of poor design.
710Nevertheless, lock-free algorithms can reduce the number of context switches, since a thread does not yield/block while waiting for a lock;
711on the other hand, a thread may busy-wait for an unbounded period.
712Finally, lock-free implementations have greater complexity and hardware dependency.
713Lock-free algorithms can be applied most easily to simple free-lists, \eg remote free-list, to allow lock-free insertion and removal from the head of a stack.
714Implementing lock-free operations for more complex data-structures (queue~\cite{Valois94}/deque~\cite{Sundell08}) is more complex.
715Michael~\cite{Michael04} and Gidenstam \etal \cite{Gidenstam05} have created lock-free variations of the Hoard allocator.
716
717
718\noindent
719====================
720
721Writing Points:
722\begin{itemize}
723\item
724Classification of benchmarks.
725\item
726Literature review of current benchmarks.
727\item
728Features and limitations.
729\item
730Literature review of current memory allocators.
731\item
732Breakdown of memory allocation techniques.
733\item
734Features and limitations.
735\end{itemize}
736
737\noindent
738====================
739
740% FIXME: cite wasik
741\cite{wasik.thesis}
742
743\section{Existing Memory Allocators}
744With dynamic allocation being an important feature of C, there are many stand-alone memory allocators that have been designed for different purposes. For this thesis, we chose 7 of the most popular and widely used memory allocators.
745
746\paragraph{dlmalloc}
747dlmalloc (FIX ME: cite allocator) is a thread-safe allocator that is single threaded and single heap. dlmalloc maintains free-lists of different sizes to store freed dynamic memory. (FIX ME: cite wasik)
748
749\paragraph{hoard}
750Hoard (FIX ME: cite allocator) is a thread-safe allocator that is multi-threaded and using a heap layer framework. It has per-thread heaps that have thread-local free-lists, and a global shared heap. (FIX ME: cite wasik)
751
752\paragraph{jemalloc}
753jemalloc (FIX ME: cite allocator) is a thread-safe allocator that uses multiple arenas. Each thread is assigned an arena. Each arena has chunks that contain contagious memory regions of same size. An arena has multiple chunks that contain regions of multiple sizes.
754
755\paragraph{ptmalloc}
756ptmalloc (FIX ME: cite allocator) is a modification of dlmalloc. It is a thread-safe multi-threaded memory allocator that uses multiple heaps. ptmalloc heap has similar design to dlmalloc's heap.
757
758\paragraph{rpmalloc}
759rpmalloc (FIX ME: cite allocator) is a thread-safe allocator that is multi-threaded and uses per-thread heap. Each heap has multiple size-classes and each size-class contains memory regions of the relevant size.
760
761\paragraph{tbb malloc}
762tbb malloc (FIX ME: cite allocator) is a thread-safe allocator that is multi-threaded and uses private heap for each thread. Each private-heap has multiple bins of different sizes. Each bin contains free regions of the same size.
763
764\paragraph{tc malloc}
765tcmalloc (FIX ME: cite allocator) is a thread-safe allocator. It uses per-thread cache to store free objects that prevents contention on shared resources in multi-threaded application. A central free-list is used to refill per-thread cache when it gets empty.
Note: See TracBrowser for help on using the repository browser.