Out of curiousity I tried KSM on the current B2G and I had some memory savings (about 20megs saved with a few apps running)
Your arguments are 100% valid however. I've also tested pksm, which seems to do something in between, hooking the mmap calls and not scanning memory (http://code.google.com/p/pksm/) The KSM scan does not seems to impact perf from a very non-scientific measurement (top ;-) If you're curious, here's the code: https://github.com/gdestuynder/releases-mozilla-central/commit/edd4c7d638639a6200703560f885f5c249aee2fb Need KSM enabled in kernel as well, then: adb shell echo 1 > /sys/kernel/mm/ksm/run Scanning happens every 20ms which is pretty often actually, but can be set to anything. Long timers will actually use more CPU as there is more to do in one run. Paged saved counted in /sys/kernel/mm/ksm/pages_shared Finally, I think it would be interesting to have something like memcow in mainline, specially so that desktop can use it as well if they go multiprocess. Guillaume On 05/04/2013 01:51 AM, Thinker K.F. Li wrote: > KSM is with higher overhead on scanning equivalent pages with MAC. > Memcow does not scan since the userspace requests COWs directly. > > KSM is working in async fashion that can not treat memory burst in a > very short time. > Memcow applys COW immediately once the userspace makes a request. > > KSM will scan all dirty pages with madvised periodically. > Memcow never do that. > t > Guillaume Destuynder <[email protected]> writes: > >> Out of curiosity, is the end result very different from KSM? (as in >> Kernel SamePage Merging) >> >> While I'm no expert with this, KSM scans for memory areas selected in >> userspace via madvise (madvise(addr, size, MADV_MERGEABLE)) then >> deduplicate it in a similar fashion and set the page as COW (it can also >> rerun the deduplication several times AFAIR, so if more identical pages >> are found overtime [that have been madvised] those would be dedup'd too) >> >> >> Guillaume >> >> On 05/03/2013 12:57 AM, Thinker K.F. Li wrote: >>> In bug 850175 and previous discussion at this list, we learnt memory >>> duplications eating out devices memory in some cases. Ting-Yuan had >>> mentioned about KSM that can merge pages with same content to avoid >>> duplication. But, it takes CPU times and is not reliable while it can >>> not merge pages in time/immediately during memory peaks. So, to >>> trigger it for OOM can not solve our issues. >>> >>> After some discussions, Ting-Yuan is trying to implement a COW >>> mechanism at userspace for sharing pages of source and target memory >>> blocks passed to memcpy. He expect to use a memdup() function to >>> replace all malloc() & memcpy() paired function calls. So, we do some >>> tricky magic to make COW applied. (I believe he will explain it >>> later.) >>> >>> Last few days, I am working on a kernel module to implement COW; call >>> it memcow. memcow provides a function that user space processes can >>> map a block of annonymous memory block to another annonymous memory >>> block, so they are sharing the same set of pages but COW; page aligned >>> of course. In another word, you can use memcow to map pages of A >>> block to the address range of B block, A and B are allocated through >>> malloc(), instead of calling memcpy. It saves a lot of memory, >>> especially for the cases like bug 850175. >>> >>> So, what do you think? >>> >>> - https://bitbucket.org/thinker/memcow >>> _______________________________________________ > _______________________________________________ dev-b2g mailing list [email protected] https://lists.mozilla.org/listinfo/dev-b2g
