I am doing to large dataset performance tests with 7.3.4b2 today and I noticed an interesting phenomenon. My shared
memorybuffers are set at 128MB. Peak postmaster usage appears to be around 90MB.
My test app performs inserts across 4 related tables, each set of 4 inserts representing a single theoretical "device"
object. I report how many "devices" I have inserted, per second, for example...
[...]
41509 devices inserted, 36/sec
[1 second later]
41544 devices inserted, 35/sec
[...]
(to be clear, 41509 devices inserted equals 166036 actual, related rows in the db)
Performance follows an odd "peak and valley" pattern. It will start out with a high insertion rate (commits are
performedafter each "device set"), then after a few thousand device sets, performance will drop to 1 device/second for
about5 seconds. Then it will slowly ramp up over the next 10 seconds to /just below/ the previous high water mark. A
fewthousand inserts later, it will drop to 1 device/second again for 5 seconds, then slowly ramp up to just below the
lasthigh water mark.
Ad infinitum.
I am wondering:
1) What am I seeing here? This is on a 4-processor machine and postmaster has a CPU all to itself, so I ruled out
processorcontention.
2) Is there more performance tuning I could perform to flatten this out, or is this just completely normal? Postmaster
neverbusts over 100MB out of the 128MB shared memory I've allocated to it, and according to <mumble mumble webpage
mumble>,this is just about perfect for shared memory settings (100 to 120% high water mark).
Thanks.
---
Clay
Cisco Systems, Inc.
claycle@cisco.com
(972) 813-5004
I've stopped 19,647 spam messages. You can too!
One month FREE spam protection at http://www.cloudmark.com/spamnetsig/}