paul@greipa.UUCP (10/18/85)
I have two 780s sitting side by side. The hardware configuration is exactly
the same: 8mb DIGITAL memory, UDA50-RA80, 4 X DUP-11, DEUNA, DMF32. The
boards are in the same slot-order in the two unibuses (unibi?). I recently
installed VMS4.1 on the second one by imaging the system pack from the first
one - thus, except for SCSNODE, all sysgen params are the same.
These two systems are parallel - they process real-time stock data. One is
a "hot backup" (which one is live varies from week to week). When the operator
started complaining that one system was slower than the other, we laughed.
Then someone ran MONITOR on both systems and one had more NULL time than the
other for the same early-morning (heavy stock trading) period.
So I wrote an integer benchmark (no FPAs, so I don't really care about FP
performance). The benchmark humps and bumps around doing useless things to
an array of 2048 longwords, 5000 times. It takes "around" 50 CPU seconds to
run.
System 1 System 2
Just VMS* 49.3 49.3
VMS + DECNet 49.6 52.2
VMS+Net+Appl** 50.1 54.3
*: this is just after the system boots. VMS is alive and I'm logged in.
**: this is after starting DECNet and the "sleep" mode of our application.
in this mode, our application takes about 10% of the CPU across six (6)
processes.
At first it bothered me that the benchmark took more CPU time on a more
heavily-loaded system: it is executing the same number of instructions
each time. Am I being charged for the context switching? I don't care,
really. My real complaint is that two (supposedly) identical systems
running identical software take different amounts of time to run the same
benchmark.
So I swapped all the CPU boards between the two. All 16 of them. It made
no difference, although I did note that the date codes on the various chips
were all 81/82 on the "slow" system and all 83 on the "fast" one.
Today I called Colorado Springs. As always, they know nothing useful. In a
few minutes I will start swapping memory and/or unibus adaptors...
Has anyone out there ever heard of this problem?
Paul Vixie
ucbvax!decwrl!greipa!paul