* Michael Galaxy (mgal...@akamai.com) wrote: > One thing to keep in mind here (despite me not having any hardware to test) > was that one of the original goals here > in the RDMA implementation was not simply raw throughput nor raw latency, > but a lack of CPU utilization in kernel > space due to the offload. While it is entirely possible that newer hardware > w/ TCP might compete, the significant > reductions in CPU usage in the TCP/IP stack were a big win at the time. > > Just something to consider while you're doing the testing........
I just noticed this thread; some random notes from a somewhat fragmented memory of this: a) Long long ago, I also tried rsocket; https://lists.gnu.org/archive/html/qemu-devel/2015-01/msg02040.html as I remember the library was quite flaky at the time. b) A lot of the complexity in the rdma migration code comes from emulating a stream to carry the migration control data and interleaving that with the actual RAM copy. I believe the original design used a separate TCP socket for the control data, and just used the RDMA for the data - that should be a lot simpler (but alas was rejected in review early on) c) I can't rememmber the last benchmarks I did; but I think I did manage to beat RDMA with multifd; but yes, multifd does eat host CPU where as RDMA barely uses a whisper. d) The 'zero-copy-send' option in migrate may well get some of that CPU time back; but if I remember we were still bottle necked on the receive side. (I can't remember if zero-copy-send worked with multifd?) e) Someone made a good suggestion (sorry can't remember who) - that the RDMA migration structure was the wrong way around - it should be the destination which initiates an RDMA read, rather than the source doing a write; then things might become a LOT simpler; you just need to send page ranges to the destination and it can pull it. That might work nicely for postcopy. Dave > - Michael > > On 5/9/24 03:58, Zheng Chuan wrote: > > Hi, Peter,Lei,Jinpu. > > > > On 2024/5/8 0:28, Peter Xu wrote: > > > On Tue, May 07, 2024 at 01:50:43AM +0000, Gonglei (Arei) wrote: > > > > Hello, > > > > > > > > > -----Original Message----- > > > > > From: Peter Xu [mailto:pet...@redhat.com] > > > > > Sent: Monday, May 6, 2024 11:18 PM > > > > > To: Gonglei (Arei) <arei.gong...@huawei.com> > > > > > Cc: Daniel P. Berrangé <berra...@redhat.com>; Markus Armbruster > > > > > <arm...@redhat.com>; Michael Galaxy <mgal...@akamai.com>; Yu Zhang > > > > > <yu.zh...@ionos.com>; Zhijian Li (Fujitsu) <lizhij...@fujitsu.com>; > > > > > Jinpu Wang > > > > > <jinpu.w...@ionos.com>; Elmar Gerdes <elmar.ger...@ionos.com>; > > > > > qemu-de...@nongnu.org; Yuval Shaia <yuval.shaia...@gmail.com>; Kevin > > > > > Wolf > > > > > <kw...@redhat.com>; Prasanna Kumar Kalever > > > > > <prasanna.kale...@redhat.com>; Cornelia Huck <coh...@redhat.com>; > > > > > Michael Roth <michael.r...@amd.com>; Prasanna Kumar Kalever > > > > > <prasanna4...@gmail.com>; integrat...@gluster.org; Paolo Bonzini > > > > > <pbonz...@redhat.com>; qemu-block@nongnu.org; de...@lists.libvirt.org; > > > > > Hanna Reitz <hre...@redhat.com>; Michael S. Tsirkin <m...@redhat.com>; > > > > > Thomas Huth <th...@redhat.com>; Eric Blake <ebl...@redhat.com>; Song > > > > > Gao <gaos...@loongson.cn>; Marc-André Lureau > > > > > <marcandre.lur...@redhat.com>; Alex Bennée <alex.ben...@linaro.org>; > > > > > Wainer dos Santos Moschetta <waine...@redhat.com>; Beraldo Leal > > > > > <bl...@redhat.com>; Pannengyuan <pannengy...@huawei.com>; > > > > > Xiexiangyou <xiexiang...@huawei.com> > > > > > Subject: Re: [PATCH-for-9.1 v2 2/3] migration: Remove RDMA protocol > > > > > handling > > > > > > > > > > On Mon, May 06, 2024 at 02:06:28AM +0000, Gonglei (Arei) wrote: > > > > > > Hi, Peter > > > > > Hey, Lei, > > > > > > > > > > Happy to see you around again after years. > > > > > > > > > Haha, me too. > > > > > > > > > > RDMA features high bandwidth, low latency (in non-blocking lossless > > > > > > network), and direct remote memory access by bypassing the CPU (As > > > > > > you > > > > > > know, CPU resources are expensive for cloud vendors, which is one of > > > > > > the reasons why we introduced offload cards.), which TCP does not > > > > > > have. > > > > > It's another cost to use offload cards, v.s. preparing more cpu > > > > > resources? > > > > > > > > > Software and hardware offload converged architecture is the way to go > > > > for all cloud vendors > > > > (Including comprehensive benefits in terms of performance, cost, > > > > security, and innovation speed), > > > > it's not just a matter of adding the resource of a DPU card. > > > > > > > > > > In some scenarios where fast live migration is needed (extremely > > > > > > short > > > > > > interruption duration and migration duration) is very useful. To > > > > > > this > > > > > > end, we have also developed RDMA support for multifd. > > > > > Will any of you upstream that work? I'm curious how intrusive would > > > > > it be > > > > > when adding it to multifd, if it can keep only 5 exported functions > > > > > like what > > > > > rdma.h does right now it'll be pretty nice. We also want to make > > > > > sure it works > > > > > with arbitrary sized loads and buffers, e.g. vfio is considering to > > > > > add IO loads to > > > > > multifd channels too. > > > > > > > > > In fact, we sent the patchset to the community in 2021. Pls see: > > > > https://urldefense.com/v3/__https://lore.kernel.org/all/20210203185906.GT2950@work-vm/T/__;!!GjvTz_vk!VfP_SV-8uRya7rBdopv8OUJkmnSi44Ktpqq1E7sr_Xcwt6zvveW51qboWOBSTChdUG1hJwfAl7HZl4NUEGc$ > > Yes, I have sent the patchset of multifd support for rdma migration by > > taking over my colleague, and also > > sorry for not keeping on this work at that time due to some reasons. > > And also I am strongly agree with Lei that the RDMA protocol has some > > special advantages against with TCP > > in some scenario, and we are indeed to use it in our product. > > > > > I wasn't aware of that for sure in the past.. > > > > > > Multifd has changed quite a bit in the last 9.0 release, that may not > > > apply > > > anymore. One thing to mention is please look at Dan's comment on possible > > > use of rsocket.h: > > > > > > https://urldefense.com/v3/__https://lore.kernel.org/all/zjjm6rcqs5eho...@redhat.com/__;!!GjvTz_vk!VfP_SV-8uRya7rBdopv8OUJkmnSi44Ktpqq1E7sr_Xcwt6zvveW51qboWOBSTChdUG1hJwfAl7HZ0CFSE-o$ > > > > > > And Jinpu did help provide an initial test result over the library: > > > > > > https://urldefense.com/v3/__https://lore.kernel.org/qemu-devel/camgffek8wiknqmouyxcathgtiem2dwocf_w7t0vmcd-i30t...@mail.gmail.com/__;!!GjvTz_vk!VfP_SV-8uRya7rBdopv8OUJkmnSi44Ktpqq1E7sr_Xcwt6zvveW51qboWOBSTChdUG1hJwfAl7HZxPNcdb4$ > > > > > > It looks like we have a chance to apply that in QEMU. > > > > > > > > > > > > One thing to note that the question here is not about a pure > > > > > performance > > > > > comparison between rdma and nics only. It's about help us make a > > > > > decision > > > > > on whether to drop rdma, iow, even if rdma performs well, the > > > > > community still > > > > > has the right to drop it if nobody can actively work and maintain it. > > > > > It's just that if nics can perform as good it's more a reason to > > > > > drop, unless > > > > > companies can help to provide good support and work together. > > > > > > > > > We are happy to provide the necessary review and maintenance work for > > > > RDMA > > > > if the community needs it. > > > > > > > > CC'ing Chuan Zheng. > > > I'm not sure whether you and Jinpu's team would like to work together and > > > provide a final solution for rdma over multifd. It could be much simpler > > > than the original 2021 proposal if the rsocket API will work out. > > > > > > Thanks, > > > > > That's a good news to see the socket abstraction for RDMA! > > When I was developed the series above, the most pain is the RDMA migration > > has no QIOChannel abstraction and i need to take a 'fake channel' > > for it which is awkward in code implementation. > > So, as far as I know, we can do this by > > i. the first thing is that we need to evaluate the rsocket is good enough > > to satisfy our QIOChannel fundamental abstraction > > ii. if it works right, then we will continue to see if it can give us > > opportunity to hide the detail of rdma protocol > > into rsocket by remove most of code in rdma.c and also some hack in > > migration main process. > > iii. implement the advanced features like multi-fd and multi-uri for rdma > > migration. > > > > Since I am not familiar with rsocket, I need some times to look at it and > > do some quick verify with rdma migration based on rsocket. > > But, yes, I am willing to involved in this refactor work and to see if we > > can make this migration feature more better:) > > > > > -- -----Open up your eyes, open up your mind, open up your code ------- / Dr. David Alan Gilbert | Running GNU/Linux | Happy \ \ dave @ treblig.org | | In Hex / \ _________________________|_____ http://www.treblig.org |_______/