Re: [PATCH 0/3] TLB flush multiple pages per IPI v4
On 04/25/2015 07:45 PM, Mel Gorman wrote: > The performance impact is documented in the changelogs but in the optimistic > case on a 4-socket machine the full series reduces interrupts from 900K > interrupts/second to 60K interrupts/second. Hello to the list, this patch have a huge (positive) performance impact on my setup. In the goal of building the best ever CDN, I run varnish web cache over very big boxes (dual xeon 12 cores, 256 Gb Ram, 24 SSd, 2*40G ethernet). Without going into varnish internal, it help to know that varnish have multiple storage backend (memory, file, etc), and that the file backend, (the one you use when you have caches drives), don't use read/write syscall but mmap. The raw performances of this server are very good : when using varnish with memory storage only, it push 80Gbps of network traffic easily. When reading/writing from/to the drives, you get 10GB/s of data. And you can do both at the same time without performance loss. Anyway, without this patch, using file storage backend and after warmup, the performance of the server was limited to a frustrating 14 Gbps. At start, varnish read from the http backend at ~ 30 Gbps, cache the data in his huge mmap, the system write it to the disk, stream it to the client, so everything looks ok. But instead of becoming quicker when the hitrate goes up (as we alread have data in the cache), it became slower and slower, to finally freeze for like 4-5 seconds every 10 sec or so. After analysis, I found out the bottleneck is the system's capacity to find free memory. If I get it correctly, when you read a "swapped out" page of a mmaped file, the kernel have to find some free memory to put the data it'll read from the drive. In my case, the disk are quick enough to handle the change almost in real time, so I've a lot of potential free memory (ie Inactive(file)). Really freeing this memory (either in direct or hard reclaim) is done relatively slowly, ie, after some tuning to avoid any direct reclaim (which was causing the freeze), I ended up having 2 kswapd (it's a bi-socket numa node) process eating 100% of cpu for ~ 14 Gbps of traffic (or ~1.5 Millions reclaims/s) After a chat with Rik van Riel and Mel Gorman, they suggest me to try this patch, and the limitation immediately jumped at 33 Gbps, which was in fact my upstream capacity, after a while I was able to achieve 60 Gbps without experiencing any issue. Even the freezing part, happening in direct reclaim mode, is a lot smoother ; on my test rig it sufficiently quick to not be seen as unavailability by my supervision (which wasn't the case before). The bad news is that after some time (like 24h) of stress testing, the performance degrade, I guess due to some kind of fragmentation. Still, the performance seems to be maintained to a higher level than the vanilla kernel. I suppose that this patch could also help a lot with database (which often mmap their data) which have to reread huge dataset frequently. Thanks a lot to Rik and Mel for the provided help, and feel free to mail me if you have question. Regards, Sébastien Wacquiez PS : the test were conducted with a 4.0.0 kernel. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH 0/3] TLB flush multiple pages per IPI v4
On 04/25/2015 07:45 PM, Mel Gorman wrote: > The performance impact is documented in the changelogs but in the optimistic > case on a 4-socket machine the full series reduces interrupts from 900K > interrupts/second to 60K interrupts/second. Hello to the list, this patch have a huge (positive) performance impact on my setup. In the goal of building the best ever CDN, I run varnish web cache over very big boxes (dual xeon 12 cores, 256 Gb Ram, 24 SSd, 2*40G ethernet). Without going into varnish internal, it help to know that varnish have multiple storage backend (memory, file, etc), and that the file backend, (the one you use when you have caches drives), don't use read/write syscall but mmap. The raw performances of this server are very good : when using varnish with memory storage only, it push 80Gbps of network traffic easily. When reading/writing from/to the drives, you get 10GB/s of data. And you can do both at the same time without performance loss. Anyway, without this patch, using file storage backend and after warmup, the performance of the server was limited to a frustrating 14 Gbps. At start, varnish read from the http backend at ~ 30 Gbps, cache the data in his huge mmap, the system write it to the disk, stream it to the client, so everything looks ok. But instead of becoming quicker when the hitrate goes up (as we alread have data in the cache), it became slower and slower, to finally freeze for like 4-5 seconds every 10 sec or so. After analysis, I found out the bottleneck is the system's capacity to find free memory. If I get it correctly, when you read a "swapped out" page of a mmaped file, the kernel have to find some free memory to put the data it'll read from the drive. In my case, the disk are quick enough to handle the change almost in real time, so I've a lot of potential free memory (ie Inactive(file)). Really freeing this memory (either in direct or hard reclaim) is done relatively slowly, ie, after some tuning to avoid any direct reclaim (which was causing the freeze), I ended up having 2 kswapd (it's a bi-socket numa node) process eating 100% of cpu for ~ 14 Gbps of traffic (or ~1.5 Millions reclaims/s) After a chat with Rik van Riel and Mel Gorman, they suggest me to try this patch, and the limitation immediately jumped at 33 Gbps, which was in fact my upstream capacity, after a while I was able to achieve 60 Gbps without experiencing any issue. Even the freezing part, happening in direct reclaim mode, is a lot smoother ; on my test rig it sufficiently quick to not be seen as unavailability by my supervision (which wasn't the case before). The bad news is that after some time (like 24h) of stress testing, the performance degrade, I guess due to some kind of fragmentation. Still, the performance seems to be maintained to a higher level than the vanilla kernel. I suppose that this patch could also help a lot with database (which often mmap their data) which have to reread huge dataset frequently. Thanks a lot to Rik and Mel for the provided help, and feel free to mail me if you have question. Regards, Sébastien Wacquiez PS : the test were conducted with a 4.0.0 kernel. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[PATCH 0/3] TLB flush multiple pages per IPI v4
The big change here is that I dropped the patch that batches TLB flushes from migration context. After V3, I realised that there are non-trivial corner cases there that deserve treatment in their own series. It did not help that I could not find a workload that was both migration and IPI intensive. The common case for IPIs during reclaim is kswapd unmapping pages which guarantees IPIs. In migration, at least some of the pages being migrated will belong to the process itself. The main issue is that migration cannot have any cached TLB entries after migration completes. Once the migration PTE is removed then writes can happen to that new page. The old TLB entry could see stale reads until it's flushed which is different to the reclaim case. This is difficult to get around. We cannot just unmap in advance because then there are no migration entries to restore and there would be minor faults post-migration. We can't batch restore the migration entries because the page lock must be held during migration or BUG_ONs get triggered. Batching TLB flushes safely requires a major rethink of how migration works so lets deal with reclaim first on its own, preferably in the context of a workload that is both migration and IPI intensive. The patch that increased the batching size was also removed because there is no advantage when TLBs are flushed before freeing the page. To increase batching we would have to alter how many pages are isolated from the LRU which would be a different patch series. Most reviewed-bys had to be dropped as the patches changed too much to preserve them. Changelog since V3 o Drop batching of TLB flush from migration o Redo how larger batching is managed o Batch TLB flushes when writable entries exist When unmapping pages it is necessary to flush the TLB. If that page was accessed by another CPU then an IPI is used to flush the remote CPU. That is a lot of IPIs if kswapd is scanning and unmapping >100K pages per second. There already is a window between when a page is unmapped and when it is TLB flushed. This series simply increases the window so multiple pages can be flushed using a single IPI. Patch 1 simply made the rest of the series easier to write as ftrace could identify all the senders of TLB flush IPIS. Patch 2 collects a list of PFNs and sends one IPI to flush them all Patch 3 tracks when there potentially are writable TLB entries that need to be batched differently The performance impact is documented in the changelogs but in the optimistic case on a 4-socket machine the full series reduces interrupts from 900K interrupts/second to 60K interrupts/second. arch/x86/Kconfig| 1 + arch/x86/include/asm/tlbflush.h | 2 + arch/x86/mm/tlb.c | 1 + include/linux/init_task.h | 8 +++ include/linux/mm_types.h| 1 + include/linux/rmap.h| 3 + include/linux/sched.h | 15 + include/trace/events/tlb.h | 3 +- init/Kconfig| 8 +++ kernel/fork.c | 5 ++ kernel/sched/core.c | 3 + mm/internal.h | 15 + mm/rmap.c | 119 +++- mm/vmscan.c | 45 ++- 14 files changed, 224 insertions(+), 5 deletions(-) -- 2.3.5 Mel Gorman (3): x86, mm: Trace when an IPI is about to be sent mm: Send one IPI per CPU to TLB flush multiple pages that were recently unmapped mm: Defer flush of writable TLB entries arch/x86/Kconfig| 1 + arch/x86/include/asm/tlbflush.h | 2 + arch/x86/mm/tlb.c | 1 + include/linux/init_task.h | 8 +++ include/linux/mm_types.h| 1 + include/linux/rmap.h| 3 + include/linux/sched.h | 15 + include/trace/events/tlb.h | 3 +- init/Kconfig| 8 +++ kernel/fork.c | 5 ++ kernel/sched/core.c | 3 + mm/internal.h | 15 + mm/rmap.c | 119 +++- mm/vmscan.c | 30 +- 14 files changed, 210 insertions(+), 4 deletions(-) -- 2.3.5 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[PATCH 0/3] TLB flush multiple pages per IPI v4
The big change here is that I dropped the patch that batches TLB flushes from migration context. After V3, I realised that there are non-trivial corner cases there that deserve treatment in their own series. It did not help that I could not find a workload that was both migration and IPI intensive. The common case for IPIs during reclaim is kswapd unmapping pages which guarantees IPIs. In migration, at least some of the pages being migrated will belong to the process itself. The main issue is that migration cannot have any cached TLB entries after migration completes. Once the migration PTE is removed then writes can happen to that new page. The old TLB entry could see stale reads until it's flushed which is different to the reclaim case. This is difficult to get around. We cannot just unmap in advance because then there are no migration entries to restore and there would be minor faults post-migration. We can't batch restore the migration entries because the page lock must be held during migration or BUG_ONs get triggered. Batching TLB flushes safely requires a major rethink of how migration works so lets deal with reclaim first on its own, preferably in the context of a workload that is both migration and IPI intensive. The patch that increased the batching size was also removed because there is no advantage when TLBs are flushed before freeing the page. To increase batching we would have to alter how many pages are isolated from the LRU which would be a different patch series. Most reviewed-bys had to be dropped as the patches changed too much to preserve them. Changelog since V3 o Drop batching of TLB flush from migration o Redo how larger batching is managed o Batch TLB flushes when writable entries exist When unmapping pages it is necessary to flush the TLB. If that page was accessed by another CPU then an IPI is used to flush the remote CPU. That is a lot of IPIs if kswapd is scanning and unmapping 100K pages per second. There already is a window between when a page is unmapped and when it is TLB flushed. This series simply increases the window so multiple pages can be flushed using a single IPI. Patch 1 simply made the rest of the series easier to write as ftrace could identify all the senders of TLB flush IPIS. Patch 2 collects a list of PFNs and sends one IPI to flush them all Patch 3 tracks when there potentially are writable TLB entries that need to be batched differently The performance impact is documented in the changelogs but in the optimistic case on a 4-socket machine the full series reduces interrupts from 900K interrupts/second to 60K interrupts/second. arch/x86/Kconfig| 1 + arch/x86/include/asm/tlbflush.h | 2 + arch/x86/mm/tlb.c | 1 + include/linux/init_task.h | 8 +++ include/linux/mm_types.h| 1 + include/linux/rmap.h| 3 + include/linux/sched.h | 15 + include/trace/events/tlb.h | 3 +- init/Kconfig| 8 +++ kernel/fork.c | 5 ++ kernel/sched/core.c | 3 + mm/internal.h | 15 + mm/rmap.c | 119 +++- mm/vmscan.c | 45 ++- 14 files changed, 224 insertions(+), 5 deletions(-) -- 2.3.5 Mel Gorman (3): x86, mm: Trace when an IPI is about to be sent mm: Send one IPI per CPU to TLB flush multiple pages that were recently unmapped mm: Defer flush of writable TLB entries arch/x86/Kconfig| 1 + arch/x86/include/asm/tlbflush.h | 2 + arch/x86/mm/tlb.c | 1 + include/linux/init_task.h | 8 +++ include/linux/mm_types.h| 1 + include/linux/rmap.h| 3 + include/linux/sched.h | 15 + include/trace/events/tlb.h | 3 +- init/Kconfig| 8 +++ kernel/fork.c | 5 ++ kernel/sched/core.c | 3 + mm/internal.h | 15 + mm/rmap.c | 119 +++- mm/vmscan.c | 30 +- 14 files changed, 210 insertions(+), 4 deletions(-) -- 2.3.5 -- To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/