On Fri, Apr 29, 2022 at 3:27 AM Aaron Lu <aaron.lu@xxxxxxxxx> wrote: > > On Fri, Apr 22, 2022 at 10:00:59AM -0700, Yang Shi wrote: > > On Thu, Apr 21, 2022 at 11:24 PM Aaron Lu <aaron.lu@xxxxxxxxx> wrote: > > > > > > On Thu, Apr 21, 2022 at 04:34:09PM +0800, ying.huang@xxxxxxxxx wrote: > > > > On Thu, 2022-04-21 at 16:17 +0800, Aaron Lu wrote: > > > > > On Thu, Apr 21, 2022 at 03:49:21PM +0800, ying.huang@xxxxxxxxx wrote: > > > > > > ... ... > > > > > > > > > For swap-in latency, we can use pmbench, which can output latency > > > > > > information. > > > > > > > > > > > > > > > > OK, I'll give pmbench a run, thanks for the suggestion. > > > > > > > > Better to construct a senario with more swapin than swapout. For > > > > example, start a memory eater, then kill it later. > > > > > > What about vm-scalability/case-swapin? > > > https://git.kernel.org/pub/scm/linux/kernel/git/wfg/vm-scalability.git/tree/case-swapin > > > > > > I think you are pretty familiar with it but still: > > > 1) it starts $nr_task processes and each mmaps $size/$nr_task area and > > > then consumes the memory, after this, it waits for a signal; > > > 2) start another process to consume $size memory to push the memory in > > > step 1) to swap device; > > > 3) kick processes in step 1) to start accessing their memory, thus > > > trigger swapins. The metric of this testcase is the swapin throughput. > > > > > > I plan to restrict the cgroup's limit to $size. > > > > > > Considering there is only one NVMe drive attached to node 0, I will run > > > the test as described before: > > > 1) bind processes to run on node 0, allocate on node 1 to test the > > > performance when reclaimer's node id is the same as swap device's. > > > 2) bind processes to run on node 1, allocate on node 0 to test the > > > performance when page's node id is the same as swap device's. > > > > > Thanks to Tim, he has found me a server that has a single Optane disk > attached to node 0. > > Let's use task0_mem0 to denote tasks bound to node 0 and memory bound > to node 0 through cgroup cpuset. And for the above swapin case: > when nr_task=1: > task0_mem0 throughput: [571652, 587158, 594316], avg=584375 -> baseline > task0_mem1 throughput: [582944, 583752, 589026], avg=585240 +0.15% > task1_mem0 throughput: [569349, 577459, 581107], avg=575971 -1.4% > task1_mem1 throughput: [564482, 570664, 571466], avg=568870 -2.6% > > task0_mem1 is slightly better than task1_mem0. > > For nr_task=8 or nr_task=16, I also gave it a run and the result is > almost the same for all 4 cases. > > > > Ying and Yang, > > > > > > Let me know what you think about the case used and the way the test is > > > conducted. > > > > Looks fine to me. To measure the latency, you could also try the below > > bpftrace script: > > > > Trying to install bpftrace on an old distro(Ubuntu 16.04) is a real > pain, I gave up... But I managed to get an old bcc installed. Using > the provided funclatency script to profile 30 seconds swap_readpage(), > there is no obvious difference from the histrogram. Thank you so much for the testing. > > So for now, from the existing results, it did't show big difference. > Theoretically, for IO device, when swapping a remote page, using the > remote swap device that is at the same node as the page can reduce the > traffic of the interlink and improve performance. I think this is the > main motivation for this code change? Yes. Given the result it seems better to keep the code as-is. > On swapin time, it's hard to say which node the task will run on anyway > so it's hard to say where to swap is beneficial. >
