Why are you messing with this code? What was wrong? The ATOMIC_CAS was the initialization of the variable running. It needs to happen. Why did you move it? The ATOMIC_GET that you first commented out and now removed actually already was just reading the variable without locks on most systems. Why did you replace it? It doesn't make *any* difference on x86_64 systems. Obviously you didn't measure the impact since there cannot have been any. On 2013-03-19 21:27, Martin Kersten wrote:
Changeset: a43f806ebd08 for MonetDB URL: http://dev.monetdb.org/hg/MonetDB?cmd=changeset;node=a43f806ebd08 Modified Files: monetdb5/mal/mal_resource.c Branch: default Log Message:
Reduce critical section in mal_resource
diffs (31 lines):
diff --git a/monetdb5/mal/mal_resource.c b/monetdb5/mal/mal_resource.c --- a/monetdb5/mal/mal_resource.c +++ b/monetdb5/mal/mal_resource.c @@ -198,7 +198,6 @@ MALresourceFairness(lng usec) if ( usec > 0 && ( (usec = GDKusec()-usec)) <= TIMESLICE ) return; threads = GDKnr_threads > 0 ? GDKnr_threads : 1; - ATOMIC_CAS_int(running, 0, threads, runningLock, "MALresourceFairness");
/* use GDKmem_cursize as MT_getrss(); is to expensive */ rss = GDKmem_cursize(); @@ -210,16 +209,14 @@ MALresourceFairness(lng usec) clk = usec / 1000;
if ( clk > DELAYUNIT ) { + ATOMIC_CAS_int(running, 0, threads, runningLock, "MALresourceFairness"); PARDEBUG mnstr_printf(GDKstdout, "#delay initial "LLFMT"n", clk); ATOMIC_DEC_int(running, runningLock, "MALresourceFairness"); - while (clk > 0) { + /* always keep one running to avoid all waiting */ + while (clk > 0 && running >= 2) { /* speed up wake up when we have memory */ if (rss < MEMORY_THRESHOLD * monet_memory) break; - /* always keep one running to avoid all waiting */ - if ( running < 2) /* dirty read of shared variable is safe here */ - //if ( (r =ATOMIC_GET_int(running, runningLock, "MALresourceFairness")) < 2) - break; delay = (unsigned int) ( ((double)DELAYUNIT * running) / threads); if (delay) { if ( delayed++ == 0){ _______________________________________________ checkin-list mailing list checkin-list@monetdb.org http://mail.monetdb.org/mailman/listinfo/checkin-list
-- Sjoerd Mullender