пн, 12 нояб. 2018 г. в 6:58, Matthew Wilcox <willy@xxxxxxxxxxxxx>: > > On Mon, Nov 12, 2018 at 12:26:10AM +0300, Timofey Titovets wrote: > > ksm by default working only on memory that added by > > madvice(). > > > > And only way get that work on other applications: > > - Use LD_PRELOAD and libraries > > - Patch kernel > > > > Lets use kernel task list in ksm_scan_thread and add logic to allow ksm > > import VMA from tasks. > > That behaviour controlled by new attribute: mode > > I try mimic hugepages attribute, so mode have two states: > > - normal - old default behaviour > > - always [new] - allow ksm to get tasks vma and try working on that. > > > > To reduce CPU load & tasklist locking time, > > ksm try import VMAs from one task per loop. > > > > So add new attribute "mode" > > Two passible values: > > - normal [default] - ksm use only madvice > > - always [new] - ksm will search vma over all processes memory and > > add it to the dedup list > > Do you have any numbers for how much difference this change makes with > various different workloads? Yep, i got some non KVM numbers, Formulas: Percentage - (pages_sharing - pages_shared)/pages_unshared Memory saved - (pages_sharing - pages_shared)*4/1024 MiB - My working laptop: 5% - ~100 MiB saved ~2GiB used Many different chrome based apps + KDE - K8s test VM: 40% - ~160 MiB saved ~920MiB used With some small running docker images - Ceph test VM: 20% - ~60MiB saved ~600MiB used With ceph mon, osd. Develop cluster servers: - K8s server backend: 72%, ~5800 MiB saved ~35.7 GiB used (With backend apps: C, java, go & etc server apps) - K8s server processing: 55%, ~2600 MiB saved ~28 GiB used (90% of load many instance of one CPU intensive application) - Ceph node: 2%, ~190 MiB saved ~11.7 GiB used (OSD only) So numbers, as always depends on the load. Thanks! - - - P.S. On recent kernels (4.19) i see BUG_ON message, that ksmd scheduled while in critical section/atomic context, not sure how to properly fix that. (If i understood correctly, i can use preempt_disable(); but that looks more like hack, not a fix). Any feedback are welcome.
