"Dirk" == Dirk Vleugels <[EMAIL PROTECTED]> writes:
Dirk> our t2k box running Generic_118833-36 experiences massive
Dirk> slowdown (upto 99% kernel cpu load) when launching several perl
Dirk> scripts. This might be related to memory starvation, but its
Dirk> hard to analyze (for me):
Dirk> top reports 16G main memory, 800m free
My first reaction upon seeing this is that the memory is fragmented
and the kernel is trying hard (probably too hard) to get large pages
for the new Perl scripts heaps.
Dirk> lockstat -s 5 -gkIW output:
Dirk> Profiling interrupt: 102240 events in 32.931 seconds (3105 events/sec)
Dirk> Count genr cuml rcnt nsec Hottest CPU+PIL Caller
Dirk>
-------------------------------------------------------------------------------
Dirk> 66840 65% ---- 0.00 3360 cpu[14]
page_geti_contig_pages
Dirk> 66553 65% ---- 0.00 3360 cpu[14] page_get_contig_pages
Dirk> 66258 65% ---- 0.00 3374 cpu[14]
page_trylock_contig_pages
Dirk> 51839 51% ---- 0.00 3404 cpu[5] page_trylock
Dirk> 32298 32% ---- 0.00 3430 cpu[5] mutex_vector_enter
Dirk> 29250 29% ---- 0.00 3315 cpu[0] page_get_freelist
Dirk> 19788 19% ---- 0.00 3112 cpu[18] thread_start
Dirk> [.....................]
Dirk> I can't make any sense from this.
Dirk> Any clues?
As I suspected, the kernel is trying to get contiguous pages to make
large pages for your applications. Once memory gets fragmented it
becomes expensive to get large pages.
As an experiment you could try running one or more of your Perl
scripts with
ppgsz -o heap=8K /path/to/script
to see if the performance problem disappears. If this performs
better, then your problem is very likely what I've described. If the
performance problems persist in this case, then the problem may be
something else.
--
Dave Marquardt
Sun Microsystems, Inc.
Austin, TX
+1 512 401-1077 (SUN internal: x64077)
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org
