Changeset 089d30c for libcfa/src/concurrency

Timestamp:

Sep 23, 2021, 1:02:46 PM (3 years ago)

Author:

Thierry Delisle <tdelisle@…>

Branches:

ADT, ast-experimental, enum, forall-pointer-decay, master, pthread-emulation, qualifiedEnum

Children:

fcd65ca

Parents:

bc4a433

Message:

Changed cpu schedulig to use moving average.

Location:

libcfa/src/concurrency

Files:

• kernel.hfa (modified) (2 diffs)
• ready_queue.cfa (modified) (7 diffs)
• thread.cfa (modified) (2 diffs)

libcfa/src/concurrency/kernel.hfa

@@ -151 +151 @@
 struct __attribute__((aligned(128))) __timestamp_t {
         volatile unsigned long long tv;
-};
-
-static inline void  ?{}(__timestamp_t & this) { this.tv = 0; }
+        volatile unsigned long long ma;
+};
+
+// Aligned timestamps which are used by the relaxed ready queue
+struct __attribute__((aligned(128))) __help_cnts_t {
+        volatile unsigned long long src;
+        volatile unsigned long long dst;
+        volatile unsigned long long tri;
+};
+
+static inline void  ?{}(__timestamp_t & this) { this.tv = 0; this.ma = 0; }
 static inline void ^?{}(__timestamp_t & this) {}
 
@@ -169 +177 @@
                 // Array of times
                 __timestamp_t * volatile tscs;
+
+                // Array of stats
+                __help_cnts_t * volatile help;
 
                 // Number of lanes (empty or not)

libcfa/src/concurrency/ready_queue.cfa

@@ -246 +246 @@
 // Cforall Ready Queue used for scheduling
 //=======================================================================
+unsigned long long moving_average(unsigned long long nval, unsigned long long oval) {
+        const unsigned long long tw = 16;
+        const unsigned long long nw = 4;
+        const unsigned long long ow = tw - nw;
+        return ((nw * nval) + (ow * oval)) / tw;
+}
+
 void ?{}(__ready_queue_t & this) with (this) {
         #if defined(USE_CPU_WORK_STEALING)
@@ -251 +258 @@
                 lanes.data = alloc( lanes.count );
                 lanes.tscs = alloc( lanes.count );
+                lanes.help = alloc( cpu_info.hthrd_count );
 
                 for( idx; (size_t)lanes.count ) {
                         (lanes.data[idx]){};
                         lanes.tscs[idx].tv = rdtscl();
+                        lanes.tscs[idx].ma = rdtscl();
+                }
+                for( idx; (size_t)cpu_info.hthrd_count ) {
+                        lanes.help[idx].src = 0;
+                        lanes.help[idx].dst = 0;
+                        lanes.help[idx].tri = 0;
                 }
         #else
                 lanes.data  = 0p;
                 lanes.tscs  = 0p;
+                lanes.help  = 0p;
                 lanes.count = 0;
         #endif
@@ -270 +285 @@
         free(lanes.data);
         free(lanes.tscs);
+        free(lanes.help);
 }
 
@@ -332 +348 @@
                 processor * const proc = kernelTLS().this_processor;
                 const int start = map.self * READYQ_SHARD_FACTOR;
+                const unsigned long long ctsc = rdtscl();
 
                 // Did we already have a help target
                 if(proc->rdq.target == -1u) {
-                        // if We don't have a
-                        unsigned long long min = ts(lanes.data[start]);
+                        unsigned long long max = 0;
                         for(i; READYQ_SHARD_FACTOR) {
-                                unsigned long long tsc = ts(lanes.data[start + i]);
-                                if(tsc < min) min = tsc;
-                        }
-                        proc->rdq.cutoff = min;
-
+                                unsigned long long tsc = moving_average(ctsc - ts(lanes.data[start + i]), lanes.tscs[start + i].ma);
+                                if(tsc > max) max = tsc;
+                        }
+                         proc->rdq.cutoff = (max + 2 * max) / 2;
                         /* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores.
                         /* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores.
 
-                        if(0 == (__tls_rand() % 10_000)) {
+                        if(0 == (__tls_rand() % 100)) {
                                 proc->rdq.target = __tls_rand() % lanes.count;
                         } else {
@@ -358 +373 @@
                 }
                 else {
-                        const unsigned long long bias = 0; //2_500_000_000;
-                        const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff;
+                        unsigned long long max = 0;
+                        for(i; READYQ_SHARD_FACTOR) {
+                                unsigned long long tsc = moving_average(ctsc - ts(lanes.data[start + i]), lanes.tscs[start + i].ma);
+                                if(tsc > max) max = tsc;
+                        }
+                        const unsigned long long cutoff = (max + 2 * max) / 2;
                         {
                                 unsigned target = proc->rdq.target;
                                 proc->rdq.target = -1u;
-                                if(lanes.tscs[target].tv < cutoff && ts(lanes.data[target]) < cutoff) {
+                                lanes.help[target].tri++;
+                                if(moving_average(ctsc - lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) {
                                         thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
                                         proc->rdq.last = target;
                                         if(t) return t;
+                                        else proc->rdq.target = -1u;
                                 }
+                                else proc->rdq.target = -1u;
                         }
 
@@ -645 +667 @@
         // Actually pop the list
         struct thread$ * thrd;
+        unsigned long long tsc_before = ts(lane);
         unsigned long long tsv;
         [thrd, tsv] = pop(lane);
@@ -658 +681 @@
         __STATS( stats.success++; )
 
-        #if defined(USE_WORK_STEALING)
+        #if defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING)
+                unsigned long long now = rdtscl();
                 lanes.tscs[w].tv = tsv;
+                lanes.tscs[w].ma = moving_average(now > tsc_before ? now - tsc_before : 0, lanes.tscs[w].ma);
         #endif
 

libcfa/src/concurrency/thread.cfa

@@ -25 +25 @@
 #include "invoke.h"
 
+uint64_t thread_rand();
+
 //-----------------------------------------------------------------------------
 // Thread ctors and dtors
@@ -41 +43 @@
         link.next = 0p;
         link.ts   = -1llu;
-        preferred = -1u;
+        preferred = thread_rand() % cl.ready_queue.lanes.count;
         last_proc = 0p;
         #if defined( __CFA_WITH_VERIFY__ )