Skip to contentSkip to conversation

Block IO: more io-cpu-affinity results

DISCUSSION Alan D. Brunelle 1 participant 2 messages Apr 15, 2008

lore
Share & tools
lore
#0 lore raw

On a 4-way IA64 box we are seeing definite improvements in overall system responsiveness w/ the patch series currently in Jens' io-cpu-affinity branch on his block IO git repository. In this microbenchmark, I peg 4 processes to 4 separate processors: 2 are doing CPU-intensive work (sqrts) and 2 are doing IO-intensive work (4KB direct reads from RAID array cache - thus limiting physical disk accesses).

There are 2 variables: whether rq_affinity is on or off for the devices under test for the IO-intensive procs, and whether the IO-intensive procs are pegged onto the same CPU as is handling IRQs for its device. The results are averaged over 4-minute runs per permutation.

When the IO-intensive procs are pegged onto the CPU that is handling IRQs for its device, we see no real difference between rq_affinity on or off:

rq=0 local=1     66.616 (M sqrt/sec)   12.312 (K ios/sec)
rq=1 local=1     66.616 (M sqrt/sec)   12.314 (K ios/sec)

Both see 66.616 million sqrts per second, and 12,300 IOs per second.

However, when we move the 2 IO-intensive threads onto CPUs that are not handling its IRQs, we see a definite improvement - both in terms of the amount of CPU-intensive work we can do (about 4%), as well as the number of IOs per second achieved (about 1%):

rq=0 local=0     61.929 (M sqrt/sec)   11.911 (K ios/sec)
rq=1 local=0     64.386 (M sqrt/sec)   12.026 (K ios/sec)

Alan

#1 lore raw
Quoted 29 lines · Alan D. Brunelle

On a 4-way IA64 box we are seeing definite improvements in overall system responsiveness w/ the patch series currently in Jens' io-cpu-affinity branch on his block IO git repository. In this microbenchmark, I peg 4 processes to 4 separate processors: 2 are doing CPU-intensive work (sqrts) and 2 are doing IO-intensive work (4KB direct reads from RAID array cache - thus limiting physical disk accesses).

There are 2 variables: whether rq_affinity is on or off for the devices under test for the IO-intensive procs, and whether the IO-intensive procs are pegged onto the same CPU as is handling IRQs for its device. The results are averaged over 4-minute runs per permutation.

When the IO-intensive procs are pegged onto the CPU that is handling IRQs for its device, we see no real difference between rq_affinity on or off:

rq=0 local=1     66.616 (M sqrt/sec)   12.312 (K ios/sec)
rq=1 local=1     66.616 (M sqrt/sec)   12.314 (K ios/sec)

Both see 66.616 million sqrts per second, and 12,300 IOs per second.

However, when we move the 2 IO-intensive threads onto CPUs that are not handling its IRQs, we see a definite improvement - both in terms of the amount of CPU-intensive work we can do (about 4%), as well as the number of IOs per second achieved (about 1%):

rq=0 local=0     61.929 (M sqrt/sec)   11.911 (K ios/sec)
rq=1 local=0     64.386 (M sqrt/sec)   12.026 (K ios/sec)
Quoted 8 lines

Alan

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

This is even more noticeable on a larger system - a 16-way IA64 box - so now 8 CPUs are doing IO-intensive and 8 are doing CPU-intensive loads.

rq=0 local=1    266.437 (M sqrt/sec)   50.018 (K ios/sec)
rq=1 local=1    266.399 (M sqrt/sec)   50.035 (K ios/sec)
rq=0 local=0    219.692 (M sqrt/sec)   39.842 (K ios/sec)
rq=1 local=0    247.406 (M sqrt/sec)   44.995 (K ios/sec)

By setting rq=1 when IOs are being remoted, we see a 12.61% improvement on the CPU-intensive processes, and 12.93% improvement for the IO-intensive loads.

However, if we remove the affinitization of the processes - just start up 16 processes (8 IO-intensive + 8 CPU-intensive), and let the scheduler associate processes w/ CPUs as normal, we see a very different picture (single run of 4 minutes per rq value):

rq=0 local=0    261.050 (M sqrt/sec)   49.147 (K ios/sec)
rq=1 local=0    264.481 (M sqrt/sec)   42.817 (K ios/sec)

Setting rq to 1 yields about a 1.31% improvement for the CPU-intensive tasks, but a 12.88% reduction in IO-intensive performance.

But that is subject to some initial placement randomness, doing ten 30-second runs, I'm seeing:

rq=0 M sqrt/sec: min=228.877, avg=240.043, max=256.925 rq=1 M sqrt/sec: min=237.202, avg=249.405, max=258.302

rq=0 K ios/sec : min= 46.198, avg= 47.760, max= 50.057 rq=1 K ios/sec : min= 38.076, avg= 41.007, max= 43.271

Which works out to a 14.14% decrease in ios/sec when RQ=1, with only a 3.90% increase in the CPU-intensive performance.

I'll need to do some work to see what's causing the problem in these latter tests...

Alan