|Pp. 165172 of the Proceedings|
Rik van Riel
The way in which a lot of this behaviour has been fixed in the Linux 2.4 kernel is described in the second part of the paper. Due to Linux 2.4 being in a code freeze period while these improvements were implemented, only known-good solutions have been integrated. A lot of the ideas used are derived from principles used in other operating systems, mostly because we have certainty that they work and a good understanding of why, making them suitable for integration into the Linux codebase during a code freeze.
The memory management in the Linux 2.2 kernel seems to be focussed on simplicity and low overhead. While this works pretty well in practice for most systems, it has some weak points left and simply falls apart under some scenarios.
Memory in Linux is unified, that is all the physical memory is on the same free list and can be allocated to any of the following memory pools on demand. Most of these pools can grow and shrink on demand. Typically most of a system's memory will be allocated to the data pages of processes and the page and buffer caches.
The page replacement of Linux 2.2 works as follows. When free memory drops below a certain threshold, the pageout daemon (kswapd) is woken up. The pageout daemon should usually be able to keep enough free memory, but if it isn't, user programs will end up calling the pageout code itself.
The main pageout loop is in the function try_to_free_pages, which starts by freeing unused slabs from the kernel memory pool. After that, it calls the following functions in a loop, asking each of them to scan a small part of their part of memory until enough memory has been freed.
Some balancing between these memory freeing function is achieved by calling them in a loop, starting of by asking each of these functions to scan a little bit of their memory, as each of these funnctions accepts a priority argument which tells them how big a percentage of their memory to scan. If not enough memory is freed in the first loop, the priority is increased and the functions are called again. The idea behind this scheme is that when one memory pool is heavily used, it will not give up its resources lightly and we'll automatically fall through to one of the other memory pools. However, this scheme relies on each of the memory pools to react in a similar way to the priority argument under different load conditions. This doesn't work out in practice because the memory pools just have fundamentally different properties to begin with.
The facts that shrink_mmap is a simple clock algorithm and relies on other functions to make process-mapped pages freeable makes it fairly unpredictable. Add to that the balancing loop in try_to_free_pages and you get a VM subsystem which is extremely sensitive to minute changes in the code and a fragile beast at its best when it comes to maintenance or (shudder) tweaking.
For Linux 2.4 a substantial development effort has gone into things like making the VM subsystem fully fine-grained for SMP systems and supporting machines with more than 1GB of RAM. Changes to the pageout code were done only in the last phase of development and are, because of that, somewhat conservative in nature and only employ known-good methods to deal with the problems that happened in the page replacement of the Linux 2.2 kernel. Before we get to the page replacement changes, however, first a short overview of the other changes in the 2.4 VM:
At the same time the memory zone for ISA DMA (0 - 16 MB physical address range) has also been split out into a separate page zone. This means larger x86 systems end up with 3 memory zones, which all need their free memory balanced so we can continue allocating kernel data structures and ISA DMA buffers. The memory zones logic is generalised enough to also work for NUMA systems.
Since the changes to the page replacement code took place after all these changes and in the (one and a half year long) code freeze period of the Linux 2.4 kernel, the changes have been kept fairly conservative. On the other hand, we have tried to fix as many of the Linux 2.2 page replacement problems as possible. Here is a short overview of the page replacement changes: they'll be described in more detail below.
The development of the page replacement changes in Linux 2.4 has been influenced by two main factors. Firstly the bad behaviours of Linux 2.2 page replacement had to be fixed, using only known-good strategies because the development of Linux 2.4 had already entered the "code freeze" state. Secondly the page replacement had to be more predictable and easier to understand than Linux 2.2 because tuning the page replacement in Linux 2.2 was deserving of the proverbial label "subtle and quick to upset". This means that only VM ideas that are well understood and have little interactions with the rest of the system were integrated. Lots of ideas were taken from other freely available operating systems and literature.
Page aging in these OSes works as follows: for each physical page we keep a counter (called age in Linux, or act_count in FreeBSD) that indicates how desirable it is to keep this page in memory. When scanning through memory for pages to evict, we increase the page age (adding a constant) whenever we find that the page was accessed and we decrease the page age (substracting a constant) whenever we find that the page wasn't accessed. When the page age (or act_count) reaches zero, the page is a candidate for eviction.
However, in some situations the LFU[Note] page aging of Linux 2.0 is known to have too much CPU overhead and adjust to changes in system load too slowly. Furthermore, research[Smaragdis, Kaplan, Wilson] has shown that recency of access is a more important criteria for page replacement than frequency.
These two problems are solved by doing exponential decline of the page age (divide by two instead of substracting a constant) whenever we find a page that wasn't accessed, resulting in page replacement which is closer to LRU[Note] than LFU. This reduces the CPU overhead of page aging drastically in some cases; however, no noticable change in swap behaviour has been observed.
Another artifact comes from the virtual address scanning. In Linux 1.2 and 2.0 the system reduces the page age of a page whenever it sees that the page hasn't been accessed from the page table which it is currently scanning, completely ignoring the fact that the page could have been accessed from other page tables. This can put a severe penalty on heavily shared pages, for example the C library.
This problem is fixed by simply not doing "downwards" aging from the virtual page scans, but only from the physical-page based scanning of the active list. If we encounter pages which are not referenced, present in the page tables but not on the active list, we simply follow the swapout path to add this page to the swap cache and the active list so we'll be able to lower the page age of this page and swap it out as soon as the page age reaches zero.
Pages which are not (yet) candidate for eviction are in process page tables, on the active list or both. Page aging as described above happens on these pages, with the function refill_inactive() balancing between scanning the page tables and scanning the active list.
When the page age on a page reaches zero, due to a combination of pageout scanning and the page not being actively used, the page is moved to the inactive_dirty list. Pages on this list are not mapped in the page tables of any process and are, or can become, reclaimable. Pages on this list are handled by the function page_launder(), which flushes the dirty pages to disk and moves the clean pages to the inactive_clean list.
Unlike the active and inactive_dirty lists, the inactive_clean list isn't global but per memory zone. The pages on these lists can be immediately reused by the page allocation code and count as free pages. These pages can also still be faulted back into where it came from, since the data is still there. In BSD this would be called the "cache" queue.
On the other hand, having lots of inactive pages to choose from when doing page eviction means you have more chances of avoiding writeout IO or doing better IO clustering. It also gives you more of a "buffer" to deal with allocations due to page faults, etc.
Both a large and a small target size for the inactive page list have their benefits. In Linux 2.4 we have chosen for a middle ground by letting the system dynamically vary the size of the inactive list depending on VM activity, with an artificial upper limit to make sure the system always preserves some aging information.
Linux 2.4 keeps a floating average of the amount of pages evicted per second and sets the target for the inactive list and the free list combined to the free target plus this average number of page steals per second. Not only does this second give us enough time to do all kinds of page flushing optimisations, it also is small enough to keep page age distribution within the system intact, allowing us to make good choices on which pages to evict and which pages to keep.
Due to the development of the page replacement changes happening in the code freeze, the system currently has a rather simple implementation of what's present in FreeBSD 4.2. As long as there are enough clean inactive pages around, we keep moving those to the inactive_clean list and never bother with syncing out the dirty pages. Note that this catches both clean pages and pages which have been written to disk by the update daemon (which commits filesystem data to disk periodically).
This means that under loads where data is seldom written we can avoid writing out dirty inactive pages most of the time, giving us much better latencies in freeing pages and letting streaming reads continue without the disk head moving away to write out data all the time. Only under loads where lots of pages are being dirtied quickly does the system suffer a bit from syncing out dirty data irregularly.
Another alternative would have been the strategy used in FreeBSD 4.3, where dirty pages get to stay in the inactive list longer than clean pages but are synced out before the clean pages are exhausted. This strategy gives more consistent pageout IO in FreeBSD during heavy write loads. However, a big factor causing the irregularities in pageout writes using the simpler strategy above may well be caused because of the huge inactive list target in FreeBSD (33It is not at all clear what this more complicated strategy would do when used on the dynamically sized inactive list on Linux 2.4, because of this Linux 2.4 uses the better understood strategy of evicting clean inactive pages first and only after those are gone start syncing the dirty ones.
Linux 2.4 is better in this respect, with the reclaim candidates neatly separated out on the inactive list. However, the inactive list could have any random size the moment VM pressure drops off. We'd like get the system in a more predictable state while the VM pressure is low. In order to achieve this, Linux 2.4 does background scanning of the pages, trying to get a sane amount of pages on the inactive list, but without scanning agressively so only truly idle pages will end up on the inactive list and the scanning overhead stays small.
Since the Linux 2.4 kernel's VM subsystem is still being tuned heavily, it is too early to come with conclusive figures on performance. However, initial results seem to indicate that Linux 2.4 generally has better performance than Linux 2.2 on the same hardware.
Reports from users indicate that performance on typical desktop machines has improved a lot, even though the tuning of the new VM has only just begun. Throughput figures for server machines seem to be better too, but that could also be attributed to the fact that the unification of the page cache and the buffer cache is complete.
One big difference between the VM in Linux 2.4 and the VM in Linux 2.2 is that the new VM is far less sensitive to subtle changes. While in Linux 2.2 a subtle change in the page flushing logic could upset page replacement, in Linux 2.4 it is possible to tweak the various aspects of the VM with predictable results and little to no side-effects in the rest of the VM.
The solid performance and relative insensitivity to subtle changes in the environment can be taken as a sign that the Linux 2.4 VM is not just a set of simple fixes for the problems experienced in Linux 2.2, but also a good base for future development.
The Linux 2.4 VM mainly contains easy to implement and obvious to verify solutions for some of the known problems Linux 2.2 suffers from. A number of issues are either too subtle to implement during the code freeze or will have too much impact on the code. The complete list of TODO items can be found on the Linux-MM page[Linux-MM]; here are the most important ones:
The author would like to thank, in no particular order: Stephen Tweedie, for taking care of memory management in Linux 1.2, 2.0 and 2.2 and also for his help with this paper; Matt Dillon, for taking the time to explain the rationale behind every little piece of the FreeBSD VM; Conectiva Inc, who employ the author to hack the Linux kernel and the wonderful crowd testers from #kernelnewbies[Kernelnewbies] and elsewhere who have helped flesh out the bugs in the Linux 2.4 VM.
This document was generated using the L aTeX2HTML translator Version 99.2beta8 (1.42)
Copyright © 1993, 1994, 1995, 1996,
Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999, Ross Moore, Mathematics Department, Macquarie University, Sydney.
The command line arguments were:
latex2html -split 0 linux24-vm.la
The translation was initiated by Rik van Riel on 2001-06-27
This paper was originally published in the
Proceedings of the FREENIX Track: 2001 USENIX Annual Technical Conference,
June 25-30, 2001, Boston, Masssachusetts, USA
Last changed: 26 July 2001 becca