One or two CPUs the pros & cons 3819

Ted MacNEIL

the redbook "effective zseries performance monitoring using resource measurement facility" gives LSPR ratios for lots of stuff.

for mixed-mode workload, 2084-332 is around 20 times the thruput of 2084-301; the 2084-302 is 1.9 times the thruput of 2084-301 for the same workload. The ratio of thruput increase to number of processors is then 1.9-2 = 95% (for 2084-302) and 20-32 = 62.5% (for 2084-332).

your number of 2084-308 with 2993-450 gives a ratio of 6.65 and for 2084-309 the ratio 3299-450 is 7.33.

the overall MIP ratio to processors of 6.65-8 is 83% and 7.33-9 is 81%. however, the incremental MIP ratio of 306-450 is 68%

from a separate source that provided 2094-7xx "SI MIPs"

for 2094-701, SI MIPs is 608 and 2094-702, SI MIPs is 1193 or an increase of 585 SI MIPs

for 2094-731, SI MIPs is 11462 and for 2094-732, SI MIPs is 11687 or an increase of 225 MIPs

the MIP ratio of 2094-732-2094-701 or 11687-608 is 19.22 ... or approx the same thruput ratio as given for 2084-332 compared to 2084-301 (i.e. the incremental 225 MIPs going from 31 to 32 results in overall effective ratio consistent with other LSPR thruput ratio comparisons).

for 2094-732, the overall SI MIP ratio to processors of 19.22-32 is 60% however the incremental SI MIP ratio of 225-608 (i.e. adding an additional processor going from 31 to 32) is 37%.

i.e. the overall increase in processor complex thruput is increasing at a slower rate than the increase in the number of processors (i.e. by 32 processors, the overall processor complex is 60% of 32 single processors). however, the incremental benefit of adding one additional processor is declining even faster (the incremental benefit of going from 31 processors to 32 processors is only 37% of a full single processor).

so possible interesting additional columns in tables for the redbook might be 1) the current thruput ratio number (to single processor) but divided by the number of processors and 2) incremental thruput divided by the thruput of a single processor.

This would show the percentage overall effective use of the complex (vis-a-vis a single processor) as the number of processors increased, plus the incremental benefit of adding one additional processor (compared to a simple single processor).

for all i know, such information is already in the redbook ... I just didn't see it as I was quickly scanning thru.

my previous post

mentioned two-processor 370 have 1.8 hardware MIP of a one processor since each machine ran at .9 cycle to allow for cross-cache chatter ... but typical thruput was in the 1.5-1.6 range because of the additional kernel overhead required to keep multiple processors coordinated.