source: doc/theses/mubeen_zulfiqar_MMath/benchmarks.tex @ 365c8dcb

\chapter{Benchmarks}

\noindent
====================

Writing Points:
\begin{itemize}
\item
Performance matrices of memory allocation.
\item
Aim of micro benchmark suite.

----- SHOULD WE GIVE IMPLEMENTATION DETAILS HERE? -----

\PAB{For the benchmarks, yes.}
\item
A complete list of benchmarks in micro benchmark suite.
\item
One detailed section for each benchmark in micro benchmark suite including:

\begin{itemize}
\item
The introduction of the benchmark.
\item
Figure.
\item
Results with popular memory allocators.
\end{itemize}

\item
Summarize performance of current memory allocators.
\end{itemize}

\noindent
====================

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Performance Matrices
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\section{Benchmarks}
There are multiple benchmarks that are built individually and evaluate different aspects of a memory allocator. But, there is not standard set of benchamrks that can be used to evaluate multiple aspects of memory allocators.

\paragraph{threadtest}
(FIX ME: cite benchmark and hoard) Each thread repeatedly allocates and then deallocates 100,000 objects. Runtime of the benchmark evaluates its efficiency.

\paragraph{shbench}
(FIX ME: cite benchmark and hoard) Each thread allocates and randomly frees a number of random-sized objects. It is a stress test that also uses runtime to determine efficiency of the allocator.

\paragraph{larson}
(FIX ME: cite benchmark and hoard) Larson simulates a server environment. Multiple threads are created where each thread allocator and free a number of objects within a size range. Some objects are passed from threads to the child threads to free. It caluculates memory operations per second as an indicator of memory allocator's performance.


\section{Performance Matrices of Memory Allocators}

When it comes to memory allocators, there are no set standards of performance. Performance of a memory allocator depends highly on the usage pattern of the application. A memory allocator that is the best performer for a certain application X might be the worst for some other application which has completely different memory usage pattern compared to the application X. It is extremely difficult to make one universally best memory allocator which will outperform every other memory allocator for every usage pattern. So, there is a lack of a set of standard benchmarks that are used to evaluate a memory allocators's performance.

If we breakdown the goals of a memory allocator, there are two basic matrices on which a memory allocator's performance is evaluated.
\begin{enumerate}
\item
Memory Overhead
\item
Speed
\end{enumerate}

\subsection{Memory Overhead}
Memory overhead is the extra memory that a memory allocator takes from OS which is not requested by the application. Ideally, an allocator should get just enough memory from OS that can fulfill application's request and should return this memory to OS as soon as applications frees it. But, allocators retain more memory compared to what application has asked for which causes memory overhead. Memory overhead can happen for various reasons.

\subsubsection{Fragmentation}
Fragmentation is one of the major reasons behind memory overhead. Fragmentation happens because of situations that are either necassary for proper functioning of the allocator such as internal memory management and book-keeping or are out of allocator's control such as application's usage pattern.

\paragraph{Internal Fragmentation}
For internal book-keeping, allocators divide raw memory given by OS into chunks, blocks, or lists that can fulfill application's requested size. Allocators use memory given by OS for creating headers, footers etc. to store information about these chunks, blocks, or lists. This increases usage of memory in-addition to the memory requested by application as the allocators need to store their book-keeping information. This extra usage of memory for allocator's own book-keeping is called Internal Fragmentation. Although it cases memory overhead but this overhead is necassary for an allocator's proper funtioning.

*** FIX ME: Insert a figure of internal fragmentation with explanation

\paragraph{External Fragmentation}
External fragmentation is the free bits of memory between or around chunks of memory that are currently in-use of the application. Segmentation in memory due to application's usage pattern causes external fragmentation. The memory which is part of external fragmentation is completely free as it is neither used by allocator's internal book-keeping nor by the application. Ideally, an allocator should return a segment of memory back to the OS as soon as application frees it. But, this is not always the case. Allocators get memory from OS in one of the two ways.

\begin{itemize}
\item
MMap: an allocator can ask OS for whole pages in mmap area. Then, the allocator segments the page internally and fulfills application's request.
\item
Heap: an allocator can ask OS for memory in heap area using system calls such as sbrk. Heap are grows downwards and shrinks upwards.
\begin{itemize}
\item
If an allocator uses mmap area, it can only return extra memory back to OS if the whole page is free i.e. no chunk on the page is in-use of the application. Even if one chunk on the whole page is currently in-use of the application, the allocator has to retain the whole page.
\item
If an allocator uses the heap area, it can only return the continous free memory at the end of the heap area that is currently in allocator's possession as heap area shrinks upwards. If there are free bits of memory in-between chunks of memory that are currently in-use of the application, the allocator can not return these free bits.

*** FIX ME: Insert a figure of above scenrio with explanation
\item
Even if the entire heap area is free except one small chunk at the end of heap area that is being used by the application, the allocator cannot return the free heap area back to the OS as it is not a continous region at the end of heap area.

*** FIX ME: Insert a figure of above scenrio with explanation

\item
Such scenerios cause external fragmentation but it is out of the allocator's control and depend on application's usage pattern.
\end{itemize}
\end{itemize}

\subsubsection{Internal Memory Management}
Allocators such as je-malloc (FIX ME: insert reference) pro-actively get some memory from the OS and divide it into chunks of certain sizes that can be used in-future to fulfill application's request. This causes memory overhead as these chunks are made before application's request. There is also the possibility that an application may not even request memory of these sizes during their whole life-time.

*** FIX ME: Insert a figure of above scenrio with explanation

Allocators such as rp-malloc (FIX ME: insert reference) maintain lists or blocks of sized memory segments that is freed by the application for future use. These lists are maintained without any guarantee that application will even request these sizes again.

Such tactics are usually used to gain speed as allocator will not have to get raw memory from OS and manage it at the time of application's request but they do cause memory overhead.

Fragmentation and managed sized chunks of free memory can lead to Heap Blowup as the allocator may not be able to use the fragments or sized free chunks of free memory to fulfill application's requests of other sizes.

\subsection{Speed}
When it comes to performance evaluation of any piece of software, its runtime is usually the first thing that is evaluated. The same is true for memory allocators but, in case of memory allocators, speed does not only mean the runtime of memory allocator's routines but there are other factors too.

\subsubsection{Runtime Speed}
Low runtime is the main goal of a memory allocator when it comes it proving its speed. Runtime is the time that it takes for a routine of memory allocator to complete its execution. As mentioned in (FIX ME: refernce to routines' list), there four basic routines that are used in memory allocation. Ideally, each routine of a memory allocator should be fast. Some memory allocator designs use pro-active measures (FIX ME: local refernce) to gain speed when allocating some memory to the application. Some memory allocators do memory allocation faster than memory freeing (FIX ME: graph refernce) while others show similar speed whether memory is allocated or freed.

\subsubsection{Memory Access Speed}
Runtime speed is not the only speed matrix in memory allocators. The memory that a memory allocator has allocated to the application also needs to be accessible as quick as possible. The application should be able to read/write allocated memory quickly. The allocation method of a memory allocator may introduce some delays when it comes to memory access speed, which is specially important in concurrent applications. Ideally, a memory allocator should allocate all memory on a cache-line to only one thread and no cache-line should be shared among multiple threads. If a memory allocator allocates memory to multple threads on a same cache line, then cache may get invalidated more frequesntly when two different threads running on two different processes will try to read/write the same memory region. On the other hand, if one cache-line is used by only one thread then the cache may get invalidated less frequently. This sharing of one cache-line among multiple threads is called false sharing (FIX ME: cite wasik).

\paragraph{Active False Sharing}
Active false sharing is the sharing of one cache-line among multiple threads that is caused by memory allocator. It happens when two threads request memory from memory allocator and the allocator allocates memory to both of them on the same cache-line. After that, if the threads are running on different processes who have their own caches and both threads start reading/writing the allocated memory simultanously, their caches will start getting invalidated every time the other thread writes something to the memory. This will cause the application to slow down as the process has to load cache much more frequently.

*** FIX ME: Insert a figure of above scenrio with explanation

\paragraph{Passive False Sharing}
Passive false sharing is the kind of false sharing which is caused by the application and not the memory allocator. The memory allocator may preservce passive false sharing in future instead of eradicating it. But, passive false sharing is initiated by the application.

\subparagraph{Program Induced Passive False Sharing}
Program induced false sharing is completely out of memory allocator's control and is purely caused by the application. When a thread in the application creates multiple objects in the dynamic area and allocator allocates memory for these objects on the same cache-line as the objects are created by the same thread. Passive false sharing will occur if this thread passes one of these objects to another thread but it retains the rest of these objects or it passes some/all of the remaining objects to some third thread(s). Now, one cache-line is shared among multiple threads but it is caused by the application and not the allocator. It is out of allocator's control and has the similar performance impact as Active False Sharing (FIX ME: cite local) if these threads, who are sharing the same cache-line, start reading/writing the given objects simultanously.

*** FIX ME: Insert a figure of above scenrio 1 with explanation

*** FIX ME: Insert a figure of above scenrio 2 with explanation

\subparagraph{Program Induced Allocator Preserved Passive False Sharing}
Program induced allocator preserved passive false sharing is another interesting case of passive false sharing. Both the application and the allocator are partially responsible for it. It starts the same as Program Induced False Sharing (FIX ME: cite local). Once, an application thread has created multiple dynamic objects on the same cache-line and ditributed these objects among multiple threads causing sharing of one cache-line among multiple threads (Program Induced Passive False Sharing). This kind of false sharing occurs when one of these threads, which got the object on the shared cache-line, frees the passed object then re-allocates another object but the allocator returns the same object (on the shared cache-line) that this thread just freed. Although, the application caused the false sharing to happen in the frst place however, to prevent furthur false sharing, the allocator should have returned the new object on some other cache-line which is only shared by the allocating thread. When it comes to performnce impact, this passive false sharing will slow down the application just like any other kind of false sharing if the threads sharing the cache-line start reading/writing the objects simultanously.


*** FIX ME: Insert a figure of above scenrio with explanation

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Micro Benchmark Suite
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\section{Micro Benchmark Suite}
The aim of micro benchmark suite is to create a set of programs that can evaluate a memory allocator based on the performance matrices described in (FIX ME: local cite). These programs can be taken as a standard to benchmark an allocator's basic goals. These programs give details of an allocator's memory overhead and speed under a certain allocation pattern. The speed of the allocator is benchmarked in different ways. Similarly, false sharing happening in an allocator is also measured in multiple ways. These benchmarks evalute the allocator under a certain allocation pattern which is configurable and can be changed using a few knobs to benchmark observe an allocator's performance under a desired allocation pattern.

Micro Benchmark Suite benchmarks an allocator's performance by allocating dynamic objects and, then, measuring specifc matrices. The benchmark suite evaluates an allocator with a certain allocation pattern. Bnechmarks have different knobs that can be used to change allocation pattern and evaluate an allocator under desired conditions. These can be set by giving commandline arguments to the benchmark on execution.

Following is the list of avalable knobs.

*** FIX ME: Add knobs items after finalize

\subsection{Memory Benchmark}
Memory benchmark measures memory overhead of an allocator. It allocates a number of dynamic objects. Then, by reading /self/proc/maps, gets the total memory that the allocator has reuested from the OS. Finally, it calculates the memory head by taking the difference between the memory the allocator has requested from the OS and the memory that program has allocated.
*** FIX ME: Insert a figure of above benchmark with description

\paragraph{Relevant Knobs}
*** FIX ME: Insert Relevant Knobs

\subsection{Speed Benchmark}
Speed benchmark calculates the runtime speed of an allocator's functions (FIX ME: cite allocator routines). It does by measuring the runtime of allocator routines in two different ways.

\subsubsection{Speed Time}
The time method does a certain amount of work by calling each routine of the allocator (FIX ME: cite allocator routines) a specific time. It calculates the total time it took to perform this workload. Then, it divides the time it took by the workload and calculates the average time taken by the allocator's routine.
*** FIX ME: Insert a figure of above benchmark with description

\paragraph{Relevant Knobs}
*** FIX ME: Insert Relevant Knobs

\subsubsection{Speed Workload}
The worload method uses the opposite approach. It calls the allocator's routines for a specific amount of time and measures how much work was done during that time. Then, similar to the time method, it divides the time by the workload done during that time and calculates the average time taken by the allocator's routine.
*** FIX ME: Insert a figure of above benchmark with description

\paragraph{Relevant Knobs}
*** FIX ME: Insert Relevant Knobs

\subsection{Cache Scratch}
Cache Scratch benchmark measures program induced allocator preserved passive false sharing (FIX ME CITE) in an allocator. It does so in two ways.

\subsubsection{Cache Scratch Time}
Cache Scratch Time allocates dynamic objects. Then, it benchmarks program induced allocator preserved passive false sharing (FIX ME CITE) in an allocator by measuring the time it takes to read/write these objects.
*** FIX ME: Insert a figure of above benchmark with description

\paragraph{Relevant Knobs}
*** FIX ME: Insert Relevant Knobs

\subsubsection{Cache Scratch Layout}
Cache Scratch Layout also allocates dynamic objects. Then, it benchmarks program induced allocator preserved passive false sharing (FIX ME CITE) by using heap addresses returned by the allocator. It calculates how many objects were allocated to different threads on the same cache line.
*** FIX ME: Insert a figure of above benchmark with description

\paragraph{Relevant Knobs}
*** FIX ME: Insert Relevant Knobs

\subsection{Cache Thrash}
Cache Thrash benchmark measures allocator induced passive false sharing (FIX ME CITE) in an allocator. It also does so in two ways.

\subsubsection{Cache Thrash Time}
Cache Thrash Time allocates dynamic objects. Then, it benchmarks allocator induced false sharing (FIX ME CITE) in an allocator by measuring the time it takes to read/write these objects.
*** FIX ME: Insert a figure of above benchmark with description

\paragraph{Relevant Knobs}
*** FIX ME: Insert Relevant Knobs

\subsubsection{Cache Thrash Layout}
Cache Thrash Layout also allocates dynamic objects. Then, it benchmarks allocator induced false sharing (FIX ME CITE) by using heap addresses returned by the allocator. It calculates how many objects were allocated to different threads on the same cache line.
*** FIX ME: Insert a figure of above benchmark with description

\paragraph{Relevant Knobs}
*** FIX ME: Insert Relevant Knobs