Index: libcfa/src/concurrency/kernel.cfa =================================================================== --- libcfa/src/concurrency/kernel.cfa (revision c34ebf2e1c30febd459338ca6bff481ddda03a67) +++ libcfa/src/concurrency/kernel.cfa (revision 69fbc6164b2feaecf3c6adb2eedc0236a52d2eb6) @@ -241,5 +241,5 @@ #if !defined(__CFA_NO_STATISTICS__) - print_stats = false; + print_stats = 0; print_halts = false; #endif @@ -281,5 +281,5 @@ #if !defined(__CFA_NO_STATISTICS__) - print_stats = false; + print_stats = 0; stats = alloc(); __init_stats( stats ); @@ -297,6 +297,6 @@ #if !defined(__CFA_NO_STATISTICS__) - if(this.print_stats) { - __print_stats( this.stats, true, this.name, (void*)&this ); + if( 0 != this.print_stats ) { + __print_stats( this.stats, this.print_stats, true, this.name, (void*)&this ); } free( this.stats ); @@ -568,6 +568,6 @@ #if !defined(__CFA_NO_STATISTICS__) __tally_stats(proc->cltr->stats, &local_stats); - if(proc->print_stats) { - __print_stats( &local_stats, true, proc->name, (void*)proc ); + if( 0 != proc->print_stats ) { + __print_stats( &local_stats, proc->print_stats, true, proc->name, (void*)proc ); } #endif Index: libcfa/src/concurrency/kernel.hfa =================================================================== --- libcfa/src/concurrency/kernel.hfa (revision c34ebf2e1c30febd459338ca6bff481ddda03a67) +++ libcfa/src/concurrency/kernel.hfa (revision 69fbc6164b2feaecf3c6adb2eedc0236a52d2eb6) @@ -106,5 +106,5 @@ #if !defined(__CFA_NO_STATISTICS__) - bool print_stats; + int print_stats; bool print_halts; #endif @@ -207,6 +207,6 @@ #if !defined(__CFA_NO_STATISTICS__) - bool print_stats; struct __stats_t * stats; + int print_stats; #endif }; @@ -229,10 +229,10 @@ #if !defined(__CFA_NO_STATISTICS__) - static inline void print_stats_at_exit( cluster & this ) { - this.print_stats = true; + static inline void print_stats_at_exit( cluster & this, int flags ) { + this.print_stats |= flags; } - static inline void print_stats_at_exit( processor & this ) { - this.print_stats = true; + static inline void print_stats_at_exit( processor & this, int flags ) { + this.print_stats |= flags; } Index: libcfa/src/concurrency/stats.cfa =================================================================== --- libcfa/src/concurrency/stats.cfa (revision c34ebf2e1c30febd459338ca6bff481ddda03a67) +++ libcfa/src/concurrency/stats.cfa (revision 69fbc6164b2feaecf3c6adb2eedc0236a52d2eb6) @@ -73,90 +73,93 @@ } - void __print_stats( struct __stats_t * stats, bool cluster, const char * name, void * id ) with( *stats ) { + void __print_stats( struct __stats_t * stats, int flags, bool cluster, const char * name, void * id ) with( *stats ) { - double push_sur = (100.0 * ((double)ready.pick.push.success) / ready.pick.push.attempt); - double pop_sur = (100.0 * ((double)ready.pick.pop .success) / ready.pick.pop .attempt); + if( flags & CFA_STATS_READY_Q ) { + double push_sur = (100.0 * ((double)ready.pick.push.success) / ready.pick.push.attempt); + double pop_sur = (100.0 * ((double)ready.pick.pop .success) / ready.pick.pop .attempt); - double push_len = ((double)ready.pick.push.attempt) / ready.pick.push.success; - double pop_len = ((double)ready.pick.pop .attempt) / ready.pick.pop .success; + double push_len = ((double)ready.pick.push.attempt) / ready.pick.push.success; + double pop_len = ((double)ready.pick.pop .attempt) / ready.pick.pop .success; - double lpush_sur = (100.0 * ((double)ready.pick.push.lsuccess) / ready.pick.push.local); - double lpop_sur = (100.0 * ((double)ready.pick.pop .lsuccess) / ready.pick.pop .local); + double lpush_sur = (100.0 * ((double)ready.pick.push.lsuccess) / ready.pick.push.local); + double lpop_sur = (100.0 * ((double)ready.pick.pop .lsuccess) / ready.pick.pop .local); - double lpush_len = ((double)ready.pick.push.local) / ready.pick.push.lsuccess; - double lpop_len = ((double)ready.pick.pop .local) / ready.pick.pop .lsuccess; + double lpush_len = ((double)ready.pick.push.local) / ready.pick.push.lsuccess; + double lpop_len = ((double)ready.pick.pop .local) / ready.pick.pop .lsuccess; + + __cfaabi_bits_print_safe( STDOUT_FILENO, + "----- %s \"%s\" (%p) - Ready Q Stats -----\n" + "- total threads run : %'15lu\n" + "- total threads scheduled: %'15lu\n" + "- push average probe len : %'18.2lf, %'18.2lf%% (%'15lu attempts)\n" + "- pop average probe len : %'18.2lf, %'18.2lf%% (%'15lu attempts)\n" + "- local push avg prb len : %'18.2lf, %'18.2lf%% (%'15lu attempts)\n" + "- local pop avg prb len : %'18.2lf, %'18.2lf%% (%'15lu attempts)\n" + "- thread migrations : %'15lu\n" + "- Idle Sleep -\n" + "-- halts : %'15lu\n" + "-- cancelled halts : %'15lu\n" + "-- schedule wake : %'15lu\n" + "-- wake on exit : %'15lu\n" + "\n" + , cluster ? "Cluster" : "Processor", name, id + , ready.pick.pop.success + , ready.pick.push.success + , push_len, push_sur, ready.pick.push.attempt + , pop_len , pop_sur , ready.pick.pop .attempt + , lpush_len, lpush_sur, ready.pick.push.local + , lpop_len , lpop_sur , ready.pick.pop .local + , ready.threads.migration + , ready.sleep.halts, ready.sleep.cancels, ready.sleep.wakes, ready.sleep.exits + ); + } #if defined(HAVE_LINUX_IO_URING_H) - double avgrdy = ((double)io.submit_q.submit_avg.rdy) / io.submit_q.submit_avg.cnt; - double avgcsm = ((double)io.submit_q.submit_avg.csm) / io.submit_q.submit_avg.cnt; - double avgavl = ((double)io.submit_q.submit_avg.avl) / io.submit_q.submit_avg.cnt; + if( flags & CFA_STATS_IO ) { + double avgrdy = ((double)io.submit_q.submit_avg.rdy) / io.submit_q.submit_avg.cnt; + double avgcsm = ((double)io.submit_q.submit_avg.csm) / io.submit_q.submit_avg.cnt; + double avgavl = ((double)io.submit_q.submit_avg.avl) / io.submit_q.submit_avg.cnt; - double lavgv = 0; - double lavgb = 0; - if(io.submit_q.look_avg.cnt != 0) { - lavgv = ((double)io.submit_q.look_avg.val ) / io.submit_q.look_avg.cnt; - lavgb = ((double)io.submit_q.look_avg.block) / io.submit_q.look_avg.cnt; + double lavgv = 0; + double lavgb = 0; + if(io.submit_q.look_avg.cnt != 0) { + lavgv = ((double)io.submit_q.look_avg.val ) / io.submit_q.look_avg.cnt; + lavgb = ((double)io.submit_q.look_avg.block) / io.submit_q.look_avg.cnt; + } + + double aavgv = 0; + double aavgb = 0; + if(io.submit_q.alloc_avg.cnt != 0) { + aavgv = ((double)io.submit_q.alloc_avg.val ) / io.submit_q.alloc_avg.cnt; + aavgb = ((double)io.submit_q.alloc_avg.block) / io.submit_q.alloc_avg.cnt; + } + + __cfaabi_bits_print_safe( STDOUT_FILENO, + "----- %s \"%s\" (%p) - I/O Stats -----\n" + "- total submit calls : %'15lu\n" + "- avg ready entries : %'18.2lf\n" + "- avg submitted entries : %'18.2lf\n" + "- avg available entries : %'18.2lf\n" + "- total ready search : %'15lu\n" + "- avg ready search len : %'18.2lf\n" + "- avg ready search block : %'18.2lf\n" + "- total alloc search : %'15lu\n" + "- avg alloc search len : %'18.2lf\n" + "- avg alloc search block : %'18.2lf\n" + "- total wait calls : %'15lu (%'lu slow, %'lu fast)\n" + "- avg completion/wait : %'18.2lf\n" + "\n" + , cluster ? "Cluster" : "Processor", name, id + , io.submit_q.submit_avg.cnt + , avgrdy, avgcsm, avgavl + , io.submit_q.look_avg.cnt + , lavgv, lavgb + , io.submit_q.alloc_avg.cnt + , aavgv, aavgb + , io.complete_q.completed_avg.slow_cnt + io.complete_q.completed_avg.fast_cnt + , io.complete_q.completed_avg.slow_cnt, io.complete_q.completed_avg.fast_cnt + , ((double)io.complete_q.completed_avg.val) / (io.complete_q.completed_avg.slow_cnt + io.complete_q.completed_avg.fast_cnt) + ); } - - double aavgv = 0; - double aavgb = 0; - if(io.submit_q.alloc_avg.cnt != 0) { - aavgv = ((double)io.submit_q.alloc_avg.val ) / io.submit_q.alloc_avg.cnt; - aavgb = ((double)io.submit_q.alloc_avg.block) / io.submit_q.alloc_avg.cnt; - } - #endif - - __cfaabi_bits_print_safe( STDOUT_FILENO, - "----- %s \"%s\" (%p) - Ready Q Stats -----\n" - "- total threads run : %'15lu\n" - "- total threads scheduled: %'15lu\n" - "- push average probe len : %'18.2lf, %'18.2lf%% (%'15lu attempts)\n" - "- pop average probe len : %'18.2lf, %'18.2lf%% (%'15lu attempts)\n" - "- local push avg prb len : %'18.2lf, %'18.2lf%% (%'15lu attempts)\n" - "- local pop avg prb len : %'18.2lf, %'18.2lf%% (%'15lu attempts)\n" - "- thread migrations : %'15lu\n" - "- Idle Sleep -\n" - "-- halts : %'15lu\n" - "-- cancelled halts : %'15lu\n" - "-- schedule wake : %'15lu\n" - "-- wake on exit : %'15lu\n" - "\n" - , cluster ? "Cluster" : "Processor", name, id - , ready.pick.pop.success - , ready.pick.push.success - , push_len, push_sur, ready.pick.push.attempt - , pop_len , pop_sur , ready.pick.pop .attempt - , lpush_len, lpush_sur, ready.pick.push.local - , lpop_len , lpop_sur , ready.pick.pop .local - , ready.threads.migration - , ready.sleep.halts, ready.sleep.cancels, ready.sleep.wakes, ready.sleep.exits - ); - #if defined(HAVE_LINUX_IO_URING_H) - __cfaabi_bits_print_safe( STDOUT_FILENO, - "----- %s \"%s\" (%p) - I/O Stats -----\n" - "- total submit calls : %'15lu\n" - "- avg ready entries : %'18.2lf\n" - "- avg submitted entries : %'18.2lf\n" - "- avg available entries : %'18.2lf\n" - "- total ready search : %'15lu\n" - "- avg ready search len : %'18.2lf\n" - "- avg ready search block : %'18.2lf\n" - "- total alloc search : %'15lu\n" - "- avg alloc search len : %'18.2lf\n" - "- avg alloc search block : %'18.2lf\n" - "- total wait calls : %'15lu (%'lu slow, %'lu fast)\n" - "- avg completion/wait : %'18.2lf\n" - "\n" - , cluster ? "Cluster" : "Processor", name, id - , io.submit_q.submit_avg.cnt - , avgrdy, avgcsm, avgavl - , io.submit_q.look_avg.cnt - , lavgv, lavgb - , io.submit_q.alloc_avg.cnt - , aavgv, aavgb - , io.complete_q.completed_avg.slow_cnt + io.complete_q.completed_avg.fast_cnt - , io.complete_q.completed_avg.slow_cnt, io.complete_q.completed_avg.fast_cnt - , ((double)io.complete_q.completed_avg.val) / (io.complete_q.completed_avg.slow_cnt + io.complete_q.completed_avg.fast_cnt) - ); #endif } Index: libcfa/src/concurrency/stats.hfa =================================================================== --- libcfa/src/concurrency/stats.hfa (revision c34ebf2e1c30febd459338ca6bff481ddda03a67) +++ libcfa/src/concurrency/stats.hfa (revision 69fbc6164b2feaecf3c6adb2eedc0236a52d2eb6) @@ -7,6 +7,13 @@ static inline void __init_stats( struct __stats_t * ) {} static inline void __tally_stats( struct __stats_t *, struct __stats_t * ) {} - static inline void __print_stats( struct __stats_t *, bool, const char *, void * ) {} + static inline void __print_stats( struct __stats_t *, int, bool, const char *, void * ) {} #else + enum { + CFA_STATS_READY_Q = 0x01, + #if defined(HAVE_LINUX_IO_URING_H) + CFA_STATS_IO = 0x02, + #endif + }; + struct __attribute__((aligned(64))) __stats_readQ_t { struct { @@ -96,5 +103,5 @@ void __init_stats ( struct __stats_t * ); void __tally_stats( struct __stats_t *, struct __stats_t * ); - void __print_stats( struct __stats_t *, bool, const char *, void * ); + void __print_stats( struct __stats_t *, int, bool, const char *, void * ); #endif