Changes in / [c86ee4c:b1a2c4a]

Location:

libcfa/src/concurrency

Files:

• invoke.h (modified) (1 diff)
• kernel.cfa (modified) (4 diffs)
• kernel.hfa (modified) (1 diff)
• kernel/startup.cfa (modified) (1 diff)
• ready_queue.cfa (modified) (1 diff)
• thread.cfa (modified) (1 diff)

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);
 
@@ -480 +470 @@
                                 #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
@@ -497 +488 @@
         #if !defined(__CFA_NO_STATISTICS__)
                 __tls_stats()->ready.threads.threads--;
+                __push_stat( __tls_stats(), __tls_stats()->ready.threads.threads, false, "Processor", this );
         #endif
 
@@ -578 +570 @@
                                 __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.
-
-                        if(0 == (__tls_rand() % 10_000)) {
-                                proc->rdq.target = __tls_rand() % lanes.count;
+                        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;
                         } else {
-                                unsigned cpu_chaos = map.start + (__tls_rand() % map.count);
-                                proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (__tls_rand() % READYQ_SHARD_FACTOR);
+                                proc->rdq.target = (cpu_chaos * READYQ_SHARD_FACTOR) + (low_chaos % 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;