On Mon, Jun 24, 2024 at 3:44 PM Huang, Ying ying.huang@intel.com wrote:
Barry Song 21cnbao@gmail.com writes:
On Fri, Jun 21, 2024 at 9:24 PM Huang, Ying ying.huang@intel.com wrote:
Barry Song 21cnbao@gmail.com writes:
On Fri, Jun 21, 2024 at 7:25 PM Ryan Roberts ryan.roberts@arm.com wrote:
On 20/06/2024 12:34, David Hildenbrand wrote:
On 20.06.24 11:04, Ryan Roberts wrote: > On 20/06/2024 01:26, Barry Song wrote: >> From: Barry Song v-songbaohua@oppo.com >> >> Both Ryan and Chris have been utilizing the small test program to aid >> in debugging and identifying issues with swap entry allocation. While >> a real or intricate workload might be more suitable for assessing the >> correctness and effectiveness of the swap allocation policy, a small >> test program presents a simpler means of understanding the problem and >> initially verifying the improvements being made. >> >> Let's endeavor to integrate it into the self-test suite. Although it >> presently only accommodates 64KB and 4KB, I'm optimistic that we can >> expand its capabilities to support multiple sizes and simulate more >> complex systems in the future as required. > > I'll try to summarize the thread with Huang Ying by suggesting this test program > is "neccessary but not sufficient" to exhaustively test the mTHP swap-out path. > I've certainly found it useful and think it would be a valuable addition to the > tree. > > That said, I'm not convinced it is a selftest; IMO a selftest should provide a > clear pass/fail result against some criteria and must be able to be run > automatically by (e.g.) a CI system.
Likely we should then consider moving other such performance-related thingies out of the selftests?
Yes, that would get my vote. But of the 4 tests you mentioned that use clock_gettime(), it looks like transhuge-stress is the only one that doesn't have a pass/fail result, so is probably the only candidate for moving.
The others either use the times as a timeout and determines failure if the action didn't occur within the timeout (e.g. ksm_tests.c) or use it to add some supplemental performance information to an otherwise functionality-oriented test.
Thank you very much, Ryan. I think you've found a better home for this tool . I will send v2, relocating it to tools/mm and adding a function to swap in either the whole mTHPs or a portion of mTHPs by "-a"(aligned swapin).
So basically, we will have
- Use MADV_PAGEPUT for rapid swap-out, putting the swap allocation code under
high exercise in a short time.
- Use MADV_DONTNEED to simulate the behavior of libc and Java heap in freeing
memory, as well as for munmap, app exits, or OOM killer scenarios. This ensures new mTHP is always generated, released or swapped out, similar to the behavior on a PC or Android phone where many applications are frequently started and terminated.
MADV_DONTNEED 64KB memory, then memset() it, this just simulates the large folio swap-in exactly, which hasn't been merged by upstream. I don't think that it's a good idea to make such kind of trick.
I disagree. This is how userspace heaps can manage memory deallocation.
Sorry, I don't understand how. Can you show some examples? Such as strace log with 64KB aligned MADV_DONTNEED?
In Java heap and memory allocators such as jemalloc and Scudo, memory is freed using the MADV_DONTNEED flag when either free() is called or garbage collection occurs. In Android, the Java heap is freed in chunks aligned to 64KB or larger. In Scudo and jemalloc, there is a configuration option to set the management granularity. This granularity is set to match the mTHP size(though the default value is 16KB in the latest Android if we don't run mTHP). Otherwise, you could end up with millions of partial unmap operations, which would severely degrade the performance of mTHP.
Imagine libc/Java functioning like a slab allocator. When kfree() is called, some pages may become completely unoccupied and can be returned to the buddy allocator. In userspace, memory is given back to the kernel in a similar manner, typically using MADV_DONTNEED. Therefore, MADV_DONTNEED is the most common memory reclamation behavior in Android, coming with free(), delete() or GC.
Imagine a system with extensive malloc, free, new, and delete operations, where objects are constantly being created and destroyed.
On the other hand, whether libc/Java use MADV_DONTNEED to free memory is not crucial, although they do. We need a method to simulate the lifecycle of applications —exiting and starting anew—on PCs or Android phones. It doesn't matter if you use MADV_DONTNEED or munmap to achieve this.
It is important to note that mTHP currently operates on a one-shot basis(after swap-out, you never get them back as mTHP as we don't support large folios swapin). For the test program, we need a method to generate new mTHPs continuously. Without this, after the initial iterations, we would be left with only folios, rendering the entire test program *pointless*.
Additionally, in the event of an application exit, munmap, or OOM killer, the amount of freed memory can be much larger than 64KB. The primary purpose of using MADV_DONTNEED is to release anonymous memory and generate new mTHP so that the iteration can continue. Otherwise, the test program becomes entirely pointless, as we only have large folios at the beginning. That is exactly why Chris has failed to find his bugs by using other small programs.
Although I still don't understand how 64KB aligned MADV_DONTNEED is used for libc/java heap or munmap in a practical way. After more thoughts, I think 64KB Aligned MADV_DONTNEED can simulate the fragmentation effect of processes exit at some degree if 64KB folios in these processes are swapped out without splitting. If you have no other practical use cases, I suggest to make it explicit with comments in program.
On the other hand, we definitely want large folios swap-in, otherwise, mTHP is just a toy to Android or similar system where more than 2/3 memory could be in swap. We do NOT want single-use mTHP.
I agree that large folios swap-in has its value at least in some situations. Whether we should take it as default behavior is another topic, we can discuss it further in the future.
Cool. Just imagine that mTHP is beneficial for systems that don't frequently use swap. However, for Android, where most memory resides in swap, mTHP acts like a one-way ticket: you end up with small folios and can't revert to large ones. This is so BAD.
- Swap in with or without the "-a" option to observe how fragments
due to swap-in and the incoming swap-in of large folios will impact swap-out fallback.
It's good to create fragmentation with swap-in. Which is more practical and future-proof. And, I believe that we can reduce large folio swap-out fallback rate without the large folio swap-in trick.
And many thanks to Chris for the suggestion on improving it within selftest, though I prefer to place it in tools/mm.
-- Best Regards, Huang, Ying
Thanks Barry