SUMMARY: mpstat - CPU contention or kernel thrashing?

Thanks to:

Kevin Buterbaugh
Chris Carrife
Joe Fletcher

for pointers.

Original question is appended below.

Chris recommended I try tying Postgres to 3 processors and see what
happened. Performance scaled down nearly linearly. I figured this took out
I/O bandwidth as an issue.

Kevin answered my question about interrupts - one CPU is handling the clock
interrupt, the other the I/O interrupts.

Joe Fletcher confirmed my feeling that the internal FC-AL drives on the V880
seem slow. I called Sun and they pointed me to a firmware upgrade for the
drives that does help.

Finally, though, I decided to play with the database - dropping the shared
buffer cache on Postgres by a factor of 9 (!!) nearly doubled performance.
Seems like its better to use the filesystem caching than Postgres? Any other
interpretations?

Thanks
Tim


-----Original Message-----
From: Tim Mohler [mailto:tamohler@nyc.rr.com]
Sent: Tuesday, February 25, 2003 4:25 PM
To: sunmanagers@sunmanagers.org
Subject: mpstat output - CPU contention or kernel thrashing?


Hi Managers -

Im running Solaris 9 on a 4-processor Sun-Fire V880 with 8GB of RAM. The
application is PostGres. We're topping out at around 260 transactions per
second (about 60% reads, 40% writes). My initial thought was that the system
was I/O bound, since the disk services times were excessive. Striping the
database boosted performance to current levels, but I'm still not satisfied.
Under a heavy load, mpstat gives output such as the following.

My questions - does the proportion of cross-calls indicate some sort of
kernel issue? Why are there two procesors with no ithr's running?

CPU minf mjf xcal  intr ithr  csw icsw migr smtx  srw syscl  usr sys  wt idl
  0 2938  78 3566   802  571 2417  405  383  175   10  4986   53  18  15  14
  1 2675  91 2646   318    2 5801 3748  399  174    9  8784   57  13  17  13
  2 2737  64 1704   341    0 2814  749  365  174   12  5471   51  14  23  12
  3 2505  60 2212  1271  894 4009 2234  347  161    3  6272   52  21  10  17
CPU minf mjf xcal  intr ithr  csw icsw migr smtx  srw syscl  usr sys  wt idl
  0 1293  96 5238  1259  834 1931  672  222   84    0  5389   87  11   1   1
  1 1241 111 3017   607    2 2335  868  229   82    0  7125   83  15   2   0
  2 1484 113 2804   538    0 2043  657  243   75    0  5264   84  15   1   0
  3 1285 102 3385  1857 1228 2249  808  245   88    0  5953   85  11   4   0
CPU minf mjf xcal  intr ithr  csw icsw migr smtx  srw syscl  usr sys  wt idl
  0 1009 113 5330  1189  766 1930  720  218  121    1  5676   80  15   4   1
  1 1027  86 2635   545    2 3676 2342  221  155    1  9077   78  13   7   2
  2 1208 102 2796   615    0 4173 2827  215  145    1  6820   80  14   5   1
  3  939  98 2818  2027 1441 4904 3648  216  142    2  8028   77  19   4   0
CPU minf mjf xcal  intr ithr  csw icsw migr smtx  srw syscl  usr sys  wt idl
  0  934 102 5379  1332  783 2227  884  171  156    0  5640   79  21   0   0
  1  959 116 3203   641    2 2380  964  201  141    0  8554   87  12   1   0
  2 1038 106 3078   768    0 2558 1027  201  158    0  5951   86  14   0   0
  3  881  96 3104  2302 1579 2185  871  181  135    0  4776   89  10   0   1

Thanks for any help,
Tim
_______________________________________________
sunmanagers mailing list
sunmanagers@sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers