On 2025-01-12, Florian Paul Schmidt <mista.tapas@xxxxxxx> wrote: > OK, I got around to playing with isolcpus. As a first attempt I added > > isolcpus=1,2,3 > > to my kernel commandline and tried > > stress -m 1 --vm-stride 16 --vm-bytes 512000000 --vm-keep -c 1 > > and it ran on core 0 maxxing it out. ... > Here's the really surprising thing though. If I run > > sudo cyclictest -m -p 95 -a 0,1,2,3 -t 4 > > then I get these enormous latencies on cores 1,2 and 3 but not on > core 0: > > # /dev/cpu_dma_latency set to 0us > policy: fifo: loadavg: 2.50 2.83 2.88 3/223 139478 > > T: 0 (139313) P:95 I:1000 C: 5881 Min: 2 Act: 3 Avg: 3 Max: 10 > T: 1 (139314) P:95 I:1500 C: 3920 Min: 1 Act: 1 Avg: 7 Max: 419 > T: 2 (139315) P:95 I:2000 C: 2940 Min: 1 Act: 1 Avg: 7 Max: 480 > T: 3 (139316) P:95 I:2500 C: 2352 Min: 1 Act: 1 Avg: 9 Max: 433 Notice the average is considerably higher on the "idle" CPUs. Perhaps you have cpufreq scaling enabled? Are you running these tests using the OSADL kernel? I assume you see the same effect when running stress(1) pinned to CPU1? ... just to be sure the boot CPU is not somehow special. (No need to boot with isolcpus since the machine is otherwise idle anyway.) taskset 0x2 stress -m 1 --vm-stride 16 --vm-bytes 512000000 --vm-keep -c 1 sudo cyclictest -m -p 95 -a 1,2,3 -t 3 John Ogness
