Changeset 50b29d9

Timestamp:

Aug 9, 2021, 4:21:44 PM (3 years ago)

Author:

Peter A. Buhr <pabuhr@…>

Branches:

ADT, ast-experimental, enum, forall-pointer-decay, jacob/cs343-translation, master, new-ast-unique-expr, pthread-emulation, qualifiedEnum

Children:

d874f59

Parents:

4f89e7b (diff), 5438e41 (diff)
Note: this is a merge changeset, the changes displayed below correspond to the merge itself.
Use the (diff) links above to see all the changes relative to each parent.

Message:

Merge branch 'master' of plg.uwaterloo.ca:software/cfa/cfa-cc

Files:

• benchmark/readyQ/cycle.cfa (modified) (1 diff)
• benchmark/readyQ/locality.cfa (modified) (2 diffs)
• benchmark/readyQ/yield.cpp (modified) (3 diffs)
• doc/theses/andrew_beach_MMath/performance.tex (modified) (4 diffs)
• libcfa/prelude/bootloader.cf (modified) (2 diffs)
• libcfa/src/concurrency/invoke.h (modified) (1 diff)
• libcfa/src/concurrency/kernel.cfa (modified) (4 diffs)
• libcfa/src/concurrency/kernel.hfa (modified) (1 diff)
• libcfa/src/concurrency/kernel/startup.cfa (modified) (1 diff)
• libcfa/src/concurrency/ready_queue.cfa (modified) (1 diff)
• libcfa/src/concurrency/thread.cfa (modified) (1 diff)

benchmark/readyQ/cycle.cfa

@@ -85 +85 @@
                 }
 
-                printf("Duration (ms)        : %'ld\n", (end - start)`dms);
+                printf("Duration (ms)        : %'lf\n", (end - start)`dms);
                 printf("Number of processors : %'d\n", nprocs);
                 printf("Number of threads    : %'d\n", tthreads);

benchmark/readyQ/locality.cfa

@@ -126 +126 @@
 // ==================================================
 // Do some work by accessing 'cnt' cells in the array
-__attribute__((noinline)) void work(MyData & data, size_t cnt, uint64_t & state) {
-        for (cnt) {
+__attribute__((noinline)) void work(MyData & data, size_t cnt_, uint64_t & state) {
+        for (cnt_) {
                 access(data, __xorshift64(state));
         }
@@ -200 +200 @@
 
         if( nspots == 0 ) { nspots = nthreads - nprocs; }
+        if( nspots == 0 ) {
+                fprintf(stderr, "--nspots must be set or --nthreads set to something bigger than --nprocs\n");
+                exit(EXIT_FAILURE);
+        }
 
         Time start, end;

benchmark/readyQ/yield.cpp

@@ -29 +29 @@
 #include "libfibre/fibre.h"
 
-FibreBarrier * barrier;
+FredBarrier * barrier;
 struct __attribute__((aligned(128))) counter_t {
         int value = 0;
@@ -66 +66 @@
 
                 const char * arg = optarg ? optarg : "";
-                size_t len = 0;
                 char * end;
                 switch(opt) {
@@ -111 +110 @@
 
                 FibreInit();
-                barrier = new FibreBarrier(nthreads + 1);
+                barrier = new FredBarrier(nthreads + 1);
                 {
                         Context::CurrCluster().addWorkers(nprocs);

doc/theses/andrew_beach_MMath/performance.tex

@@ -160 +160 @@
 % Match All     &   3719462155 &   43294042 &   3223004977 &  1286054154 &   623887874 \\
 % Match None    &   4971630929 &   55311709 &   9481225467 &  1310251289 &   623752624 \\
+%
+% run-algol-04-a
+% --------------
+% Raise Empty   & 0.0 & 0.0 &  3250260945 & 0.0 & 0.0 \\
+% Raise D'tor   & 0.0 & 0.0 & 29017675113 & N/A & N/A \\
+% Raise Finally & 0.0 & 0.0 &         N/A & 0.0 & 0.0 \\
+% Raise Other   & 0.0 & 0.0 & 24411823773 & 0.0 & 0.0 \\
+% Cross Handler & 0.0 & 0.0 &      769334 & 0.0 & 0.0 \\
+% Cross Finally & 0.0 & N/A &         N/A & 0.0 & 0.0 \\
+% Match All     & 0.0 & 0.0 &  3254283504 & 0.0 & 0.0 \\
+% Match None    & 0.0 & 0.0 &  9476060146 & 0.0 & 0.0 \\
+
 \begin{tabular}{|l|c c c c c|}
 \hline
@@ -196 +208 @@
 % Match All     &   3189512499 &   39124453 &   2667795989 &  1525889031 &   733785613 \\
 % Match None    &   4094675477 &   48749857 &   7850618572 &  1566713577 &   733478963 \\
+%
+% run-plg7a-04-a
+% --------------
+% 0.0 are unfilled.
+% Raise Empty   & 0.0 & 0.0 &  2770781479 & 0.0 & 0.0 \\
+% Raise D'tor   & 0.0 & 0.0 & 23530084907 & N/A & N/A \\
+% Raise Finally & 0.0 & 0.0 &         N/A & 0.0 & 0.0 \\
+% Raise Other   & 0.0 & 0.0 & 23816827982 & 0.0 & 0.0 \\
+% Cross Handler & 0.0 & 0.0 &     1422188 & 0.0 & 0.0 \\
+% Cross Finally & 0.0 & N/A &         N/A & 0.0 & 0.0 \\
+% Match All     & 0.0 & 0.0 &  2671989778 & 0.0 & 0.0 \\
+% Match None    & 0.0 & 0.0 &  7829059869 & 0.0 & 0.0 \\
 
 % PLG7A (in seconds)
@@ -202 +226 @@
               & \CFA (Terminate) & \CFA (Resume) & \Cpp & Java & Python \\
 \hline
-% Raise Empty   & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 \\
-% Raise D'tor   & 0.0 & 0.0 & 0.0 & N/A & N/A \\
-% Raise Finally & 0.0 & 0.0 & N/A & 0.0 & 0.0 \\
-% Raise Other   & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 \\
-% Cross Handler & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 \\
-% Cross Finally & 0.0 & N/A & N/A & 0.0 & 0.0 \\
-% Match All     & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 \\
-% Match None    & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 \\
 Raise Empty   & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 \\
 Raise D'tor   & 0.0 & 0.0 & 0.0 & N/A & N/A \\
@@ -231 +247 @@
 
 This means that while \CFA does not actually keep up with Python in every
-case it is no worse than roughly half the speed of \Cpp. This is good
+case it is usually no worse than roughly half the speed of \Cpp. This is good
 enough for the prototyping purposes of the project.
 
-One difference not shown is that optimizations in \CFA is very fragile.
-The \CFA compiler uses gcc as part of its complation process and the version
-of gcc could change the speed of some of the benchmarks by 10 times or more.
-Similar changes to g++ for the \Cpp benchmarks had no significant changes.
-Because of the connection between gcc and g++; this suggests it is not the
-optimizations that are changing but how the optimizer is detecting if the
-optimizations can be applied. So the optimizations are always applied in
-g++, but only newer versions of gcc can detect that they can be applied in
-the more complex \CFA code.
+The test case where \CFA falls short is Raise Other, the case where the
+stack is unwound including a bunch of non-matching handlers.
+This slowdown seems to come from missing optimizations,
+the results above came from gcc/g++ 10 (gcc as \CFA backend or g++ for \Cpp)
+but the results change if they are run in gcc/g++ 9 instead.
+Importantly, there is a huge slowdown in \Cpp's results bringing that brings
+\CFA's performace back in that roughly half speed area. However many other
+\CFA benchmarks increase their run-time by a similar amount falling far
+behind their \Cpp counter-parts.
+
+This suggests that the performance issue in Raise Other is just an
+optimization not being applied. Later versions of gcc may be able to
+optimize this case further, at least down to the half of \Cpp mark.
+A \CFA compiler that directly produced assembly could do even better as it
+would not have to work across some of \CFA's current abstractions, like
+the try terminate function.
 
 Resumption exception handling is also incredibly fast. Often an order of

libcfa/prelude/bootloader.cf

@@ -3 +3 @@
 char ** cfa_args_argv;
 char ** cfa_args_envp;
+int cfa_main_returned = 0;
 
 int main(int argc, char* argv[], char* envp[]) {
@@ -8 +9 @@
         cfa_args_argv = argv;
         cfa_args_envp = envp;
-        return invoke_main(argc, argv, envp);
+        int ret = invoke_main(argc, argv, envp);
+        cfa_main_returned = 1;
+        return ret;
 }

libcfa/src/concurrency/invoke.h

@@ -170 +170 @@
                 bool corctx_flag;
 
+                int last_cpu;
+
                 //SKULLDUGGERY errno is not save in the thread data structure because returnToKernel appears to be the only function to require saving and restoring it
 

libcfa/src/concurrency/kernel.cfa

@@ -394 +394 @@
         __builtin_prefetch( thrd_dst->context.SP );
 
+        int curr = __kernel_getcpu();
+        if(thrd_dst->last_cpu != curr) {
+                int64_t l = thrd_dst->last_cpu;
+                int64_t c = curr;
+                int64_t v = (l << 32) | c;
+                __push_stat( __tls_stats(), v, false, "Processor", this );
+        }
+
+        thrd_dst->last_cpu = curr;
+
         __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);
 
@@ -470 +480 @@
                                 #if !defined(__CFA_NO_STATISTICS__)
                                         __tls_stats()->ready.threads.threads++;
-                                        __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
                                 #endif
                                 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
@@ -488 +497 @@
         #if !defined(__CFA_NO_STATISTICS__)
                 __tls_stats()->ready.threads.threads--;
-                __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
         #endif
 
@@ -570 +578 @@
                                 __tls_stats()->ready.threads.extunpark++;
                         }
-                        __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", kernelTLS().this_processor );
                 }
                 else {
                         __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
                         __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
-                        __push_stat( cl->stats, cl->stats->ready.threads.threads, true, "Cluster", cl );
                 }
         #endif

libcfa/src/concurrency/kernel.hfa

@@ -67 +67 @@
                 unsigned target;
                 unsigned last;
+                unsigned cnt;
                 unsigned long long int cutoff;
         } rdq;

libcfa/src/concurrency/kernel/startup.cfa

@@ -478 +478 @@
         state = Start;
         self_cor{ info };
+        last_cpu = __kernel_getcpu();
         curr_cor = &self_cor;
         curr_cluster = mainCluster;

libcfa/src/concurrency/ready_queue.cfa

@@ -345 +345 @@
                         /* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores.
                         /* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores.
-                        uint64_t chaos = __tls_rand();
-                        uint64_t high_chaos = (chaos >> 32);
-                        uint64_t  mid_chaos = (chaos >> 16) & 0xffff;
-                        uint64_t  low_chaos = chaos & 0xffff;
-
-                        unsigned me = map.self;
-                        unsigned cpu_chaos = map.start + (mid_chaos % map.count);
-                        bool global = cpu_chaos == me;
-
-                        if(global) {
-                                proc->rdq.target = high_chaos % lanes.count;
+
+                        if(0 == (__tls_rand() % 10_000)) {
+                                proc->rdq.target = __tls_rand() % lanes.count;
                         } else {
-                                proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (low_chaos % READYQ_SHARD_FACTOR);
+                                unsigned cpu_chaos = map.start + (__tls_rand() % map.count);
+                                proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (__tls_rand() % READYQ_SHARD_FACTOR);
                                 /* paranoid */ verify(proc->rdq.target >= (map.start * READYQ_SHARD_FACTOR));
                                 /* paranoid */ verify(proc->rdq.target <  ((map.start + map.count) * READYQ_SHARD_FACTOR));

libcfa/src/concurrency/thread.cfa

@@ -34 +34 @@
         preempted = __NO_PREEMPTION;
         corctx_flag = false;
+        last_cpu = __kernel_getcpu();
         curr_cor = &self_cor;
         self_mon.owner = &this;