From: Christophe de Dinechin <dinechin@redhat.com> To: Paolo Bonzini <pbonzini@redhat.com> Cc: KVM list <kvm@vger.kernel.org>, "Michael S. Tsirkin" <mst@redhat.com>, Marcelo Tosatti <mtosatti@redhat.com>, Xiao Guangrong <xiaoguangrong@tencent.com>, "Dr. David Alan Gilbert" <dgilbert@redhat.com>, Peter Xu <peterx@redhat.com>, qemu-devel@nongnu.org, Juan Quintela <quintela@redhat.com>, Wei Wang <wei.w.wang@intel.com>, "Emilio G. Cota" <cota@braap.org>, Xiao Guangrong <guangrong.xiao@gmail.com>, jiang.biao2@zte.com.cn Subject: Re: [PATCH v3 2/5] util: introduce threaded workqueue Date: Tue, 4 Dec 2018 16:49:29 +0100 [thread overview] Message-ID: <3A78700E-A32B-42DC-90FA-F1501D434608@redhat.com> (raw) In-Reply-To: <3a17b878-9a1c-7cdc-0250-187e82e2faf3@redhat.com> > On 27 Nov 2018, at 14:51, Paolo Bonzini <pbonzini@redhat.com> wrote: > > On 27/11/18 13:49, Christophe de Dinechin wrote: >> So this is not really >> helping. Also, the ThreadLocal structure itself is not necessarily aligned >> within struct Threads. Therefore, it’s possible that “requests” for example >> could be on the same cache line as request_fill_bitmap if planets align >> the wrong way. > > I think this is a bit exaggerated. Hence my “if planets align the wrong way” :-) But I understand that my wording came out too strong. My apologies. I think the fix is to align ThreadLocal as well. > Linux and QEMU's own qht work just fine with compile-time directives. Wouldn’t it work fine without any compile-time directive at all? Alignment is just a performance optimization. > >> In order to mitigate these effects, I would group the data that the user >> writes and the data that the thread writes, i.e. reorder declarations, >> put request_fill_bitmap and request_valid_ev together, and try >> to put them in the same cache line so that only one cache line is invalidated >> from within mark_request_valid instead of two. >> >> Then you end up with a single alignment directive instead of 4, to >> separate requests from completions. > > Yeah, I agree with this. > >> That being said, I’m not sure why you use a bitmap here. What is the >> expected benefit relative to atomic lists (which would also make it really >> lock-free)? >> > > I don't think lock-free lists are easier. Bitmaps smaller than 64 > elements are both faster and easier to manage. I believe that this is only true if you use a linked list for both freelist management and for thread notification (i.e. to replace the bitmaps). However, if you use an atomic list only for the free list, and keep bitmaps for signaling, then performance is at least equal, often better. Plus you get the added benefit of having a thread-safe API, i.e. something that is truly lock-free. I did a small experiment to test / prove this. Last commit on branch: https://github.com/c3d/recorder/commits/181122-xiao_guangdong_introduce-threaded-workqueue Take with a grain of salt, microbenchmarks are always suspect ;-) The code in “thread_test.c” includes Xiao’s code with two variations, plus some testing code lifted from the flight recorder library. 1. The FREE_LIST variation (sl_test) is what I would like to propose. 2. The BITMAP variation (bm_test) is the baseline 3. The DOUBLE_LIST variation (ll_test) is the slow double-list approach To run it, you need to do “make opt-test”, then run “test_script” which outputs a CSV file. The summary of my findings testing on a ThinkPad, a Xeon machine and a MacBook is here: https://imgur.com/a/4HmbB9K Overall, the proposed approach: - makes the API thread safe and lock free, addressing the one drawback that Xiao was mentioning. - delivers up to 30% more requests on the Macbook, while being “within noise” (sometimes marginally better) for the other two. I suspect an optimization opportunity found by clang, because the Macbook delivers really high numbers. - spends less time blocking when all threads are busy, which accounts for the higher number of client loops. If you think that makes sense, then either Xiao can adapt the code from the branch above, or I can send a follow-up patch. Thanks Christophe
WARNING: multiple messages have this Message-ID (diff)
From: Christophe de Dinechin <dinechin@redhat.com> To: Paolo Bonzini <pbonzini@redhat.com> Cc: Xiao Guangrong <guangrong.xiao@gmail.com>, "Michael S. Tsirkin" <mst@redhat.com>, Marcelo Tosatti <mtosatti@redhat.com>, qemu-devel@nongnu.org, KVM list <kvm@vger.kernel.org>, "Dr. David Alan Gilbert" <dgilbert@redhat.com>, Peter Xu <peterx@redhat.com>, Wei Wang <wei.w.wang@intel.com>, jiang.biao2@zte.com.cn, Eric Blake <eblake@redhat.com>, Juan Quintela <quintela@redhat.com>, "Emilio G. Cota" <cota@braap.org>, Xiao Guangrong <xiaoguangrong@tencent.com> Subject: Re: [Qemu-devel] [PATCH v3 2/5] util: introduce threaded workqueue Date: Tue, 4 Dec 2018 16:49:29 +0100 [thread overview] Message-ID: <3A78700E-A32B-42DC-90FA-F1501D434608@redhat.com> (raw) In-Reply-To: <3a17b878-9a1c-7cdc-0250-187e82e2faf3@redhat.com> > On 27 Nov 2018, at 14:51, Paolo Bonzini <pbonzini@redhat.com> wrote: > > On 27/11/18 13:49, Christophe de Dinechin wrote: >> So this is not really >> helping. Also, the ThreadLocal structure itself is not necessarily aligned >> within struct Threads. Therefore, it’s possible that “requests” for example >> could be on the same cache line as request_fill_bitmap if planets align >> the wrong way. > > I think this is a bit exaggerated. Hence my “if planets align the wrong way” :-) But I understand that my wording came out too strong. My apologies. I think the fix is to align ThreadLocal as well. > Linux and QEMU's own qht work just fine with compile-time directives. Wouldn’t it work fine without any compile-time directive at all? Alignment is just a performance optimization. > >> In order to mitigate these effects, I would group the data that the user >> writes and the data that the thread writes, i.e. reorder declarations, >> put request_fill_bitmap and request_valid_ev together, and try >> to put them in the same cache line so that only one cache line is invalidated >> from within mark_request_valid instead of two. >> >> Then you end up with a single alignment directive instead of 4, to >> separate requests from completions. > > Yeah, I agree with this. > >> That being said, I’m not sure why you use a bitmap here. What is the >> expected benefit relative to atomic lists (which would also make it really >> lock-free)? >> > > I don't think lock-free lists are easier. Bitmaps smaller than 64 > elements are both faster and easier to manage. I believe that this is only true if you use a linked list for both freelist management and for thread notification (i.e. to replace the bitmaps). However, if you use an atomic list only for the free list, and keep bitmaps for signaling, then performance is at least equal, often better. Plus you get the added benefit of having a thread-safe API, i.e. something that is truly lock-free. I did a small experiment to test / prove this. Last commit on branch: https://github.com/c3d/recorder/commits/181122-xiao_guangdong_introduce-threaded-workqueue Take with a grain of salt, microbenchmarks are always suspect ;-) The code in “thread_test.c” includes Xiao’s code with two variations, plus some testing code lifted from the flight recorder library. 1. The FREE_LIST variation (sl_test) is what I would like to propose. 2. The BITMAP variation (bm_test) is the baseline 3. The DOUBLE_LIST variation (ll_test) is the slow double-list approach To run it, you need to do “make opt-test”, then run “test_script” which outputs a CSV file. The summary of my findings testing on a ThinkPad, a Xeon machine and a MacBook is here: https://imgur.com/a/4HmbB9K Overall, the proposed approach: - makes the API thread safe and lock free, addressing the one drawback that Xiao was mentioning. - delivers up to 30% more requests on the Macbook, while being “within noise” (sometimes marginally better) for the other two. I suspect an optimization opportunity found by clang, because the Macbook delivers really high numbers. - spends less time blocking when all threads are busy, which accounts for the higher number of client loops. If you think that makes sense, then either Xiao can adapt the code from the branch above, or I can send a follow-up patch. Thanks Christophe
next prev parent reply other threads:[~2018-12-04 15:49 UTC|newest] Thread overview: 70+ messages / expand[flat|nested] mbox.gz Atom feed top 2018-11-22 7:20 [PATCH v3 0/5] migration: improve multithreads guangrong.xiao 2018-11-22 7:20 ` [Qemu-devel] " guangrong.xiao 2018-11-22 7:20 ` [PATCH v3 1/5] bitops: introduce change_bit_atomic guangrong.xiao 2018-11-22 7:20 ` [Qemu-devel] " guangrong.xiao 2018-11-23 10:23 ` Dr. David Alan Gilbert 2018-11-23 10:23 ` [Qemu-devel] " Dr. David Alan Gilbert 2018-11-28 9:35 ` Juan Quintela 2018-11-28 9:35 ` [Qemu-devel] " Juan Quintela 2018-11-22 7:20 ` [PATCH v3 2/5] util: introduce threaded workqueue guangrong.xiao 2018-11-22 7:20 ` [Qemu-devel] " guangrong.xiao 2018-11-23 11:02 ` Dr. David Alan Gilbert 2018-11-23 11:02 ` [Qemu-devel] " Dr. David Alan Gilbert 2018-11-26 7:57 ` Xiao Guangrong 2018-11-26 7:57 ` [Qemu-devel] " Xiao Guangrong 2018-11-26 10:56 ` Dr. David Alan Gilbert 2018-11-26 10:56 ` [Qemu-devel] " Dr. David Alan Gilbert 2018-11-27 7:17 ` Xiao Guangrong 2018-11-27 7:17 ` [Qemu-devel] " Xiao Guangrong 2018-11-26 18:55 ` Emilio G. Cota 2018-11-26 18:55 ` [Qemu-devel] " Emilio G. Cota 2018-11-27 8:30 ` Xiao Guangrong 2018-11-27 8:30 ` [Qemu-devel] " Xiao Guangrong 2018-11-24 0:12 ` Emilio G. Cota 2018-11-24 0:12 ` [Qemu-devel] " Emilio G. Cota 2018-11-26 8:06 ` Xiao Guangrong 2018-11-26 8:06 ` [Qemu-devel] " Xiao Guangrong 2018-11-26 18:49 ` Emilio G. Cota 2018-11-26 18:49 ` [Qemu-devel] " Emilio G. Cota 2018-11-27 8:29 ` Xiao Guangrong 2018-11-27 8:29 ` [Qemu-devel] " Xiao Guangrong 2018-11-24 0:17 ` Emilio G. Cota 2018-11-24 0:17 ` [Qemu-devel] " Emilio G. Cota 2018-11-26 8:18 ` Xiao Guangrong 2018-11-26 8:18 ` [Qemu-devel] " Xiao Guangrong 2018-11-26 10:28 ` Paolo Bonzini 2018-11-26 10:28 ` [Qemu-devel] " Paolo Bonzini 2018-11-27 8:31 ` Xiao Guangrong 2018-11-27 8:31 ` [Qemu-devel] " Xiao Guangrong 2018-11-27 12:49 ` Christophe de Dinechin 2018-11-27 12:49 ` [Qemu-devel] " Christophe de Dinechin 2018-11-27 13:51 ` Paolo Bonzini 2018-11-27 13:51 ` [Qemu-devel] " Paolo Bonzini 2018-12-04 15:49 ` Christophe de Dinechin [this message] 2018-12-04 15:49 ` Christophe de Dinechin 2018-12-04 17:16 ` Paolo Bonzini 2018-12-04 17:16 ` [Qemu-devel] " Paolo Bonzini 2018-12-10 3:23 ` Xiao Guangrong 2018-12-10 3:23 ` [Qemu-devel] " Xiao Guangrong 2018-11-27 17:39 ` Emilio G. Cota 2018-11-27 17:39 ` [Qemu-devel] " Emilio G. Cota 2018-11-28 8:55 ` Xiao Guangrong 2018-11-28 8:55 ` [Qemu-devel] " Xiao Guangrong 2018-11-22 7:20 ` [PATCH v3 3/5] migration: use threaded workqueue for compression guangrong.xiao 2018-11-22 7:20 ` [Qemu-devel] " guangrong.xiao 2018-11-23 18:17 ` Dr. David Alan Gilbert 2018-11-23 18:17 ` [Qemu-devel] " Dr. David Alan Gilbert 2018-11-23 18:22 ` Paolo Bonzini 2018-11-23 18:22 ` [Qemu-devel] " Paolo Bonzini 2018-11-23 18:29 ` Dr. David Alan Gilbert 2018-11-23 18:29 ` [Qemu-devel] " Dr. David Alan Gilbert 2018-11-26 8:00 ` Xiao Guangrong 2018-11-26 8:00 ` [Qemu-devel] " Xiao Guangrong 2018-11-22 7:20 ` [PATCH v3 4/5] migration: use threaded workqueue for decompression guangrong.xiao 2018-11-22 7:20 ` [Qemu-devel] " guangrong.xiao 2018-11-22 7:20 ` [PATCH v3 5/5] tests: add threaded-workqueue-bench guangrong.xiao 2018-11-22 7:20 ` [Qemu-devel] " guangrong.xiao 2018-11-22 21:25 ` [PATCH v3 0/5] migration: improve multithreads no-reply 2018-11-22 21:25 ` [Qemu-devel] " no-reply 2018-11-22 21:35 ` no-reply 2018-11-22 21:35 ` [Qemu-devel] " no-reply
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=3A78700E-A32B-42DC-90FA-F1501D434608@redhat.com \ --to=dinechin@redhat.com \ --cc=cota@braap.org \ --cc=dgilbert@redhat.com \ --cc=guangrong.xiao@gmail.com \ --cc=jiang.biao2@zte.com.cn \ --cc=kvm@vger.kernel.org \ --cc=mst@redhat.com \ --cc=mtosatti@redhat.com \ --cc=pbonzini@redhat.com \ --cc=peterx@redhat.com \ --cc=qemu-devel@nongnu.org \ --cc=quintela@redhat.com \ --cc=wei.w.wang@intel.com \ --cc=xiaoguangrong@tencent.com \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.