On Wed, Nov 27, 2024 at 12:26 AM Johannes Weiner hannes@cmpxchg.org wrote:
On Tue, Nov 26, 2024 at 06:57:19PM -0800, Yuanchu Xie wrote:
This patch series provides workingset reporting of user pages in lruvecs, of which coldness can be tracked by accessed bits and fd references. However, the concept of workingset applies generically to all types of memory, which could be kernel slab caches, discardable userspace caches (databases), or CXL.mem. Therefore, data sources might come from slab shrinkers, device drivers, or the userspace. Another interesting idea might be hugepage workingset, so that we can measure the proportion of hugepages backing cold memory. However, with architectures like arm, there may be too many hugepage sizes leading to a combinatorial explosion when exporting stats to the userspace. Nonetheless, the kernel should provide a set of workingset interfaces that is generic enough to accommodate the various use cases, and extensible to potential future use cases.
Doesn't DAMON already provide this information?
Yuanchu might be able to answer this question a lot better than I do, since he studied DAMON and tried to leverage it in our fleet.
My impression is that there are some fundamental differences in access detection and accounting mechanisms between the two, i.e., sampling vs scanning-based detection and non-lruvec vs lruvec-based accounting.
CCing SJ.
Use cases
Job scheduling On overcommitted hosts, workingset information improves efficiency and reliability by allowing the job scheduler to have better stats on the exact memory requirements of each job. This can manifest in efficiency by landing more jobs on the same host or NUMA node. On the other hand, the job scheduler can also ensure each node has a sufficient amount of memory and does not enter direct reclaim or the kernel OOM path. With workingset information and job priority, the userspace OOM killing or proactive reclaim policy can kick in before the system is under memory pressure. If the job shape is very different from the machine shape, knowing the workingset per-node can also help inform page allocation policies.
Proactive reclaim Workingset information allows the a container manager to proactively reclaim memory while not impacting a job's performance. While PSI may provide a reactive measure of when a proactive reclaim has reclaimed too much, workingset reporting allows the policy to be more accurate and flexible.
I'm not sure about more accurate.
Agreed. This is a (very) poor argument, unless there are facts to back this up.
Access frequency is only half the picture. Whether you need to keep memory with a given frequency resident depends on the speed of the backing device.
Along a similar line, we also need to consider use cases that don't involve backing storage, e.g., far memory (remote node). More details below.
There is memory compression; there is swap on flash; swap on crappy flash; swapfiles that share IOPS with co-located filesystems. There is zswap+writeback, where avg refault speed can vary dramatically.
You can of course offload much more to a fast zswap backend than to a swapfile on a struggling flashdrive, with comparable app performance.
So I think you'd be hard pressed to achieve a high level of accuracy in the usecases you list without taking the (often highly dynamic) cost of paging / memory transfer into account.
There is a more detailed discussion of this in a paper we wrote on proactive reclaim/offloading - in 2.5 Hardware Heterogeneity:
https://www.cs.cmu.edu/~dskarlat/publications/tmo_asplos22.pdf
Ballooning (similar to proactive reclaim) The last patch of the series extends the virtio-balloon device to report the guest workingset. Balloon policies benefit from workingset to more precisely determine the size of the memory balloon. On end-user devices where memory is scarce and overcommitted, the balloon sizing in multiple VMs running on the same device can be orchestrated with workingset reports from each one. On the server side, workingset reporting allows the balloon controller to inflate the balloon without causing too much file cache to be reclaimed in the guest.
Promotion/Demotion If different mechanisms are used for promition and demotion, workingset information can help connect the two and avoid pages being migrated back and forth. For example, given a promotion hot page threshold defined in reaccess distance of N seconds (promote pages accessed more often than every N seconds). The threshold N should be set so that ~80% (e.g.) of pages on the fast memory node passes the threshold. This calculation can be done with workingset reports. To be directly useful for promotion policies, the workingset report interfaces need to be extended to report hotness and gather hotness information from the devices[1].
[1] https://www.opencompute.org/documents/ocp-cms-hotness-tracking-requirements-...
Sysfs and Cgroup Interfaces
The interfaces are detailed in the patches that introduce them. The main idea here is we break down the workingset per-node per-memcg into time intervals (ms), e.g.
1000 anon=137368 file=24530 20000 anon=34342 file=0 30000 anon=353232 file=333608 40000 anon=407198 file=206052 9223372036854775807 anon=4925624 file=892892
Implementation
The reporting of user pages is based off of MGLRU, and therefore requires CONFIG_LRU_GEN=y. We would benefit from more MGLRU generations for a more fine-grained workingset report, but we can already gather a lot of data with just four generations. The workingset reporting mechanism is gated behind CONFIG_WORKINGSET_REPORT, and the aging thread is behind CONFIG_WORKINGSET_REPORT_AGING.
Benchmarks
Ghait Ouled Amar Ben Cheikh has implemented a simple policy and ran Linux compile and redis benchmarks from openbenchmarking.org. The policy and runner is referred to as WMO (Workload Memory Optimization). The results were based on v3 of the series, but v4 doesn't change the core of the working set reporting and just adds the ballooning counterpart.
The timed Linux kernel compilation benchmark shows improvements in peak memory usage with a policy of "swap out all bytes colder than 10 seconds every 40 seconds". A swapfile is configured on SSD.
peak memory usage (with WMO): 4982.61328 MiB peak memory usage (control): 9569.1367 MiB peak memory reduction: 47.9%
Benchmark | Experimental |Control | Experimental_Std_Dev | Control_Std_Dev Timed Linux Kernel Compilation - allmodconfig (sec) | 708.486 (95.91%) | 679.499 (100%) | 0.6% | 0.1%
Seconds, fewer is better
You can do this with a recent (>2018) upstream kernel and ~100 lines of python [1]. It also works on both LRU implementations.
[1] https://github.com/facebookincubator/senpai
We use this approach in virtually the entire Meta fleet, to offload unneeded memory, estimate available capacity for job scheduling, plan future capacity needs, and provide accurate memory usage feedback to application developers.
It works over a wide variety of CPU and storage configurations with no specific tuning.
How would Senpai work for use cases that don't have local storage, i.e., all memory is mapped by either the fast or the slow tier? (>95% memory usage in our fleet is mapped and local storage for non-storage servers is only scratch space.)
My current understanding is that its approach would not be able to form a feedback loop because there are currently no refaults from the slow tier (because it's also mapped), and that's where I think this proposal or something similar can help.
Also this proposal reports histograms, not scalars. So in theory, userspace can see the projections of its potential actions, rather than solely rely on trial and error. Of course, this needs to be backed with data. So yes, some comparisons from real-world use cases would be very helpful to demonstrate the value of this proposal.