Direct compact when a high-order allocation fails
in [Kernel]
|
Prev: __weak vs ifdef
Next: kdb,panic,debug_core: Allow the debug core to receive a panic before smp_send_stop()
From: KAMEZAWA Hiroyuki on 24 Mar 2010 21:10 On Wed, 24 Mar 2010 13:48:16 -0700 Andrew Morton <akpm(a)linux-foundation.org> wrote: > On Tue, 23 Mar 2010 12:25:45 +0000 > Mel Gorman <mel(a)csn.ul.ie> wrote: > > + /* > > + * We will not stall if the necessary conditions are not met for > > + * migration but direct reclaim seems to account stalls similarly > > + */ > > + count_vm_event(COMPACTSTALL); > > + > > + /* Compact each zone in the list */ > > + for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, > > + nodemask) { > > Will all of this code play nicely with memory hotplug? > If your concern is a race with memory hotplug, I have no concern about that because memory hotplug makes a range of pages as "not for use" before starting. If your concern is "code sharing", shared codes between memory hotplug and compaction is "migrate_pages()". Other parts are independent from each other. IIUC. Memory Hotremove does 1. select a range for removal [start ....end) 2. mark free pages as "not for use" by migrate_type 3. move all used pages to other range. 4. Finally, all pages in the range will be "not for use" Compaction does 1. select a target order 2. move some free pages to private list 3. move some used pages to pages in private list. 4. free pages. So, techniques to isolate freed pages is different. I think it's from their purpose. "freed pages" by compaction is - for use - a chunk of page from anywhere is ok. but "freed pages" by memory unplug is - not for use - a chunk of page should be in specified range. For using memory hotplug's code for compaction, we have to specify "not for use" range. It will make low order compaction innefficient and it seems not easy to find the best range for compaction. For compaction, logic used in memory hotplug is too big hummer, I guess. Thanks, -Kame -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: KAMEZAWA Hiroyuki on 25 Mar 2010 06:00 On Thu, 25 Mar 2010 09:48:26 +0000 Mel Gorman <mel(a)csn.ul.ie> wrote: > > In that case, compact_finished() can't > > find there is a free chunk and do more work. How about using a function like > > free_pcppages_bulk(zone, pcp->batch, pcp); > > to bypass pcp list and freeing pages at once ? > > > > I think you mean to drain the PCP lists while compaction is happening > but is it justified? It's potentially a lot of IPI calls just to check > if compaction can finish a little earlier. If the pages on the PCP lists > are making that much of a difference to high-order page availability, it > implies that the zone is pretty full and it's likely that compaction was > avoided and we direct reclaimed. > Ah, sorry for my short word again. I mean draining "local" pcp list because a thread which run direct-compaction freed pages. IPI is not necessary and overkill. Thanks, -Kame -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Mel Gorman on 25 Mar 2010 06:20 On Thu, Mar 25, 2010 at 06:50:21PM +0900, KAMEZAWA Hiroyuki wrote: > On Thu, 25 Mar 2010 09:48:26 +0000 > Mel Gorman <mel(a)csn.ul.ie> wrote: > > > > In that case, compact_finished() can't > > > find there is a free chunk and do more work. How about using a function like > > > free_pcppages_bulk(zone, pcp->batch, pcp); > > > to bypass pcp list and freeing pages at once ? > > > > > > > I think you mean to drain the PCP lists while compaction is happening > > but is it justified? It's potentially a lot of IPI calls just to check > > if compaction can finish a little earlier. If the pages on the PCP lists > > are making that much of a difference to high-order page availability, it > > implies that the zone is pretty full and it's likely that compaction was > > avoided and we direct reclaimed. > > > Ah, sorry for my short word again. I mean draining "local" pcp list because > a thread which run direct-compaction freed pages. IPI is not necessary and > overkill. > Ah, I see now. There are two places that pages get freed. release_freepages() at the end of compaction when it's too late for compact_finished() to be helped and within migration itself. Migration frees with either free_page() or more commonly put_page() with put_page() being the most frequently used. As free_page() is called on failure to migrate (rare), there is little help in changing it and I'd rather not modify how put_page() works. I could add a variant of drain_local_pages() that drains just the local PCP of a given zone before compact_finished() is called. The cost would be a doubling of the number of times zone->lock is taken to do the drain. Is it justified? It seems overkill to me to take the zone->lock just in case compaction can finish a little earlier. It feels like it would be adding a guaranteed cost for a potential saving. -- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: KOSAKI Motohiro on 25 Mar 2010 07:10 > > Hmm..Hmmm........... > > > > Today, I've reviewed this patch and [11/11] carefully twice. but It is harder to ack. > > > > This patch seems to assume page compaction is faster than direct > > reclaim. but it often doesn't, because dropping useless page cache is very > > lightweight operation, > > Two points with that; > > 1. It's very hard to know in advance how often direct reclaim of clean page > cache would be enough to satisfy the allocation. Yeah, This is main reason why I'd suggest tightly integrate vmscan and compaction. > 2. Even if it was faster to discard page cache, it's not necessarily > faster when the cost of reading that page cache back-in is taken into > account _If_ this is useful page, you are right. but please remember, In typical case the system have lots no longer used pages. > > Lumpy reclaim tries to avoid dumping useful page cache but it is perfectly > possible for hot data to be discarded because it happened to be located > near cold data. Yeah, I fully agree. > It's impossible to know in general how much unnecessary IO > takes place as a result of lumpy reclaim because it depends heavily on the > system-state when lumpy reclaim starts. I think this is explained why vmscan and compaction shouldn't be separated. Yes, Only vmscan can know it. > > but page compaction makes a lot of memcpy (i.e. cpu cache > > pollution). IOW this patch is focusing to hugepage allocation very aggressively, but > > it seems not enough care to reduce typical workload damage. > > > > What typical workload is making aggressive use of high order > allocations? Typically when such a user is found, effort is spent on > finding alternatives to high-orders as opposed to worrying about the cost > of allocating them. There was a focus on huge page allocation because it > was the most useful test case that was likely to be encountered in practice. > > I can adjust the allocation levels to some other value but it's not typical > for a system to make very aggressive use of other orders. I could have it > use random orders but also is not very typical. If this compaction is trigged only order-9 allocation, I don't oppose it at all. Also PAGE_ALLOC_COSTLY_ORDER is probably acceptable. I agree huge page allocation made lots trouble. but low order and the system have lots no longer used page case, your logic is worse than current. I worry about it. My point is, We have to consider to disard useful cached pages and to discard no longer accessed pages. latter is nearly zero cost. please don't consider page discard itself is bad, it is correct page life cycle. To protest discard useless cached page can makes reduce IO throughput. > > > At first, I would like to clarify current reclaim corner case and how > > vmscan should do at this mail. > > > > Now we have Lumpy reclaim. It is very excellent solution for externa > > fragmentation. > > In some situations, it can grind a system to trash for a time. What is far > more likely is to be dealing with a machine with no swap - something that > is common in clusters. In this case, lumpy is a lot less likely to succeed > unless the machine is very quiet. It's just not going to find the contiguous > page cache it needs to discard and anonymous pages get in the way. > > > but unfortunately it have lots corner case. > > > > Viewpoint 1. Unnecessary IO > > > > isolate_pages() for lumpy reclaim frequently grab very young page. it is often > > still dirty. then, pageout() is called much. > > > > Unfortunately, page size grained io is _very_ inefficient. it can makes lots disk > > seek and kill disk io bandwidth. > > > > Page-based IO like this has also been reported as being a problem for some > filesystems. When this happens, lumpy reclaim potentially stalls for a long > time waiting for the dirty data to be flushed by a flusher thread. Compaction > does not suffer from the same problem. > > > Viewpoint 2. Unevictable pages > > > > isolate_pages() for lumpy reclaim can pick up unevictable page. it is obviously > > undroppable. so if the zone have plenty mlocked pages (it is not rare case on > > server use case), lumpy reclaim can become very useless. > > > > Also true. Potentially, compaction can deal with unevictable pages but it's > not done in this series as it's significant enough as it is and useful in > its current form. > > > Viewpoint 3. GFP_ATOMIC allocation failure > > > > Obviously lumpy reclaim can't help GFP_ATOMIC issue. > > > > Also true although right now, it's not possible to compact for GFP_ATOMIC > either. I think it could be done on some cases but I didn't try for it. > High-order GFP_ATOMIC allocations are still something we simply try and > avoid rather than deal with within the page allocator. > > > Viewpoint 4. reclaim latency > > > > reclaim latency directly affect page allocation latency. so if lumpy reclaim with > > much pageout io is slow (often it is), it affect page allocation latency and can > > reduce end user experience. > > > > Also true. When allocation huge pages on a normal desktop for example, > it scan stall the machine for a number of seconds while reclaim kicks > in. > > With direct compaction, this does not happen to anywhere near the same > degree. There are still some stalls because as huge pages get allocated, > free memory drops until pages have to be reclaimed anyway. The effects > are a lot less prononced and the operation finishes a lot faster. > > > I really hope that auto page migration help to solve above issue. but sadly this > > patch seems doesn't. > > > > How do you figure? I think it goes a long way to mitigating the worst of > the problems you laid out above. Both lumpy reclaim and page comaction have some advantage and some disadvantage. However we already have lumpy reclaim. I hope you rememver we are attacking very narrowing corner case. we have to consider to reduce the downside of compaction at first priority. Not only big benefit but also big downside seems no good. So, I'd suggest either way 1) no change caller place, but invoke compaction at very limited situation, or 2) invoke compaction at only lumpy reclaim unfit situation My last mail, I proposed about (2). but you seems got bad impression. then, now I propsed (1). I mean we will _start_ to treat the compaction is for hugepage allocation assistance feature, not generic allocation change. btw, I hope drop or improve patch 11/11 ;-) > > Honestly, I think this patch was very impressive and useful at 2-3 years ago. > > because 1) we didn't have lumpy reclaim 2) we didn't have sane reclaim bail out. > > then, old vmscan is very heavyweight and inefficient operation for high order reclaim. > > therefore the downside of adding this page migration is hidden relatively. but... > > > > We have to make an effort to reduce reclaim latency, not adding new latency source. > > I recognise that reclaim latency has been reduced but there is a wall. If it is a wall, we have to fix this! :) > The cost of reading the data back in will always be there and on > swapless systems, it might simply be impossible for lumpy reclaim to do > what it needs. Well, I didn't and don't think the compaction is useless. I haven't saied the compaction is useless. I've talked about how to avoid downside mess. > > Instead, I would recommend tightly integrate page-compaction and lumpy reclaim. > > I mean 1) reusing lumpy reclaim's neighbor pfn page pickking up logic > > There are a number of difficulties with this. I'm not saying it's impossible, > but the win is not very clear-cut and there are some disadvantages. > > One, there would have to be exceptions for kswapd in the path because it > really should continue reclaiming. The reclaim path is already very dense > and this would add significant compliexity to that path. > > The second difficulty is that the migration and free block selection > algorithm becomes a lot harder, more expensive and identifying the exit > conditions presents a significant difficultly. Right now, the selection is > based on linear scans with straight-forward selection and the exit condition > is simply when the scanners meet. With the migration scanner based on LRU, > significant care would have to be taken to ensure that appropriate free blocks > were chosen to migrate to so that we didn't "migrate from" a block in one > pass and "migrate to" in another (the reason why I went with linear scans > in the first place). Identifying when the zone has been compacted and should > just stop is no longer as straight-forward either. You'd have to track what > blocks had been operated on in the past which is potentially a lot of state. To > maintain this state, an unknown number structures would have to be allocated > which may re-enter the allocator presenting its own class of problems. > > Third, right now it's very easy to identify when compaction is not going > to work in advance - simply check the watermarks and make a calculation > based on fragmentation. With a combined reclaim/compaction step, these > type of checks would need to be made continually - potentially > increasing the latency of reclaim albeit very slightly. > > Lastly, with this series, there is very little difference between direct > compaction and proc-triggered compaction. They share the same code paths > and all that differs is the exit conditions. If it was integrated into > reclaim, it becomes a lot less straight-forward to share the code. > > > 2) do page > > migration instead pageout when the page is some condition (example active or dirty > > or referenced or swapbacked). > > > > Right now, it is identifed when pageout should happen instead of page > migration. It's known before compaction starts if it's likely to be > successful or not. > patch 11/11 says, it's known likely to be successfull or not, but not exactly. I think you and I don't have so big different analisys about current behavior. I feel I merely pesimistic rather than you. > > This patch seems shoot me! /me die. R.I.P. ;-) > > That seems a bit dramatic. Your alternative proposal has some significant > difficulties and is likely to be very complicated. Also, there is nothing > to say that this mechanism could not be integrated with lumpy reclaim over > time once it was shown that useless migration was going on or latencies were > increased for some workload. > > This patch seems like a far more rational starting point to me than adding > more complexity to reclaim at the outset. > > > btw please don't use 'hugeadm --set-recommended-min_free_kbytes' at testing. > > It's somewhat important for the type of stress tests I do for huge page > allocation. Without it, fragmentation avoidance has trouble and the > results become a lot less repeatable. > > > To evaluate a case of free memory starvation is very important for this patch > > series, I think. I slightly doubt this patch might invoke useless compaction > > in such case. > > > > I can drop the min_free_kbytes change but the likely result will be that > allocation success rates will simply be lower. The calculations on > whether compaction should be used or not are based on watermarks which > adjust to the value of min_free_kbytes. Then, should we need min_free_kbytes auto adjustment trick? > > At bottom line, the explict compaction via /proc can be merged soon, I think. > > but this auto compaction logic seems need more discussion. > > My concern would be that the compaction paths would then be used very > rarely in practice and we'd get no data on how direct compaction should > be done. Agree almost. Again, I think this patch is attacking corner case issue. then, I don't hope this will makes new corner case. I don't think your approach is perfectly broken. But please remember, now compaction might makes very large lru shuffling in compaction failure case. It mean vmscan might discard very wrong pages. I have big worry about it. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: KOSAKI Motohiro on 25 Mar 2010 07:30
> > Viewpoint 5. end user surprising > > > > lumpy reclaim can makes swap-out even though the system have lots free > > memory. end users very surprised it and they can think it is bug. > > > > Also, this swap activity easyly confuse that an administrator decide when > > install more memory into the system. > > > > Compaction in this case is a lot less surprising. If there is enough free > memory, compaction will trigger automatically without any reclaim. I fully agree this. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ |