Subject: 10gR1 SLES-9 Possible parallelism issue 2007-08-20 - By Michael Taylor
Back
Suse-Oracle gurus,
During a 64-bit conversion I completed on a new quad socket dual core DL 580
running SLES-9 (See http://LES-9.ora-code.com) SP3 and 10.1.0.5 RAC over Polyserve CFS this weekend, I observed
that only 4 CPUs were doing anything (other than being idle) during the CPU
intensive recompile portion. I have done probably over a hundred database
upgrade on multiprocessor machines, and generally find that recompiling
invalidated objects works best (completes in the shortest time) when using a
parallelism of 4 times the number of CPU cores. Since we have 8 cores on
the 580s, I used a parallelism of 32. I could not complete the conversion
in RAC mode, so I could only use one machine.
@(protected)
exec utl_recomp.recomp_parallel(32);
Usually when I kick off a recompile in this manner, I see 32 job workers kicked
off at the database level, and when watching top, I can see 32 processes (on
linux, make the xterm window longer) running at high CPU utilization.
Are system setting or kernel parameter (sysctl.conf) that
might be only allowing 4 processes to run on CPU simultaneously? I am not
concerned as much about the recompile time as that when we turn on the
64-bit database in production, only 4 CPU cores will do any work, and the other
4 will sit around idle. The sysadmins indicate they built a kernel with ?make
?j 32? and saw at least 12 instances of gcc running
concurrently.
I am hoping this is not a Xeon limitation. Last time I did a 64-bit
conversion on an Opteron based HP DL 585 running SLES-9 (See http://LES-9.ora-code.com), I saw fantastic
performance on the recompile, and 32 oracle processes were seen via top as
described above.
Thanks,
-Michael Taylor
__ ____ ____ ____ ____ ____ ____ ____ ____ ____ ____ ____ ____ ____ ____ ______
__ __
Take the Internet to Go: Yahoo!Go puts the Internet in your pocket: mail, news,
photos & more.
http://mobile.yahoo.com/go?refer=1GNXIC
|
|
