Hi Peter,
On Mon, Apr 8, 2024 at 6:18 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
>
> On Mon, Apr 08, 2024 at 04:07:20PM +0200, Jinpu Wang wrote:
> > Hi Peter,
>
> Jinpu,
>
> Thanks for joining the discussion.
>
> >
> > On Tue, Apr 2, 2024 at 11:24 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
> > >
> > > On Mon, Apr 01, 2024 at 11:26:25PM +0200, Yu Zhang wrote:
> > > > Hello Peter und Zhjian,
> > > >
> > > > Thank you so much for letting me know about this. I'm also a bit surprised at
> > > > the plan for deprecating the RDMA migration subsystem.
> > >
> > > It's not too late, since it looks like we do have users not yet notified
> > > from this, we'll redo the deprecation procedure even if it'll be the final
> > > plan, and it'll be 2 releases after this.
> > >
> > > >
> > > > > IMHO it's more important to know whether there are still users and whether
> > > > > they would still like to see it around.
> > > >
> > > > > I admit RDMA migration was lack of testing(unit/CI test), which led to the a few
> > > > > obvious bugs being noticed too late.
> > > >
> > > > Yes, we are a user of this subsystem. I was unaware of the lack of test coverage
> > > > for this part. As soon as 8.2 was released, I saw that many of the
> > > > migration test
> > > > cases failed and came to realize that there might be a bug between 8.1
> > > > and 8.2, but
> > > > was unable to confirm and report it quickly to you.
> > > >
> > > > The maintenance of this part could be too costly or difficult from
> > > > your point of view.
> > >
> > > It may or may not be too costly, it's just that we need real users of RDMA
> > > taking some care of it. Having it broken easily for >1 releases definitely
> > > is a sign of lack of users. It is an implication to the community that we
> > > should consider dropping some features so that we can get the best use of
> > > the community resources for the things that may have a broader audience.
> > >
> > > One thing majorly missing is a RDMA tester to guard all the merges to not
> > > break RDMA paths, hopefully in CI. That should not rely on RDMA hardwares
> > > but just to sanity check the migration+rdma code running all fine. RDMA
> > > taught us the lesson so we're requesting CI coverage for all other new
> > > features that will be merged at least for migration subsystem, so that we
> > > plan to not merge anything that is not covered by CI unless extremely
> > > necessary in the future.
> > >
> > > For sure CI is not the only missing part, but I'd say we should start with
> > > it, then someone should also take care of the code even if only in
> > > maintenance mode (no new feature to add on top).
> > >
> > > >
> > > > My concern is, this plan will forces a few QEMU users (not sure how
> > > > many) like us
> > > > either to stick to the RDMA migration by using an increasingly older
> > > > version of QEMU,
> > > > or to abandon the currently used RDMA migration.
> > >
> > > RDMA doesn't get new features anyway, if there's specific use case for RDMA
> > > migrations, would it work if such a scenario uses the old binary? Is it
> > > possible to switch to the TCP protocol with some good NICs?
> > We have used rdma migration with HCA from Nvidia for years, our
> > experience is RDMA migration works better than tcp (over ipoib).
>
> Please bare with me, as I know little on rdma stuff.
>
> I'm actually pretty confused (and since a long time ago..) on why we need
> to operation with rdma contexts when ipoib seems to provide all the tcp
> layers. I meant, can it work with the current "tcp:" protocol with ipoib
> even if there's rdma/ib hardwares underneath? Is it because of performance
> improvements so that we must use a separate path comparing to generic
> "tcp:" protocol here?
using rdma protocol with ib verbs , we can leverage the full benefit of RDMA by
talking directly to NIC which bypasses the kernel overhead, less cpu
utilization and better performance.
While IPoIB is more for compatibility to applications using tcp, but
can't get full benefit of RDMA. When you have mix generation of IB
devices, there are performance issue on IPoIB, we've seen 40G HCA can
only reach 2 Gb/s on IPoIB, but with raw RDMA can reach full line
speed.
I just run a simple iperf3 test via ipoib and ib_send_bw on same hosts:
iperf 3.9
Linux ps404a-3 5.15.137-pserver #5.15.137-6~deb11 SMP Thu Jan 4
07:19:34 UTC 2024 x86_64
-----------------------------------------------------------
Server listening on 5201
-----------------------------------------------------------
Time: Tue, 09 Apr 2024 06:55:02 GMT
Accepted connection from 2a02:247f:401:4:2:0:b:3, port 41130
Cookie: cer2hexlldrowclq6izh7gbg5toviffqbcwt
TCP MSS: 0 (default)
[ 5] local 2a02:247f:401:4:2:0:a:3 port 5201 connected to
2a02:247f:401:4:2:0:b:3 port 41136
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks, omitting
0 seconds, 10 second test, tos 0
[ ID] Interval Transfer Bitrate
[ 5] 0.00-1.00 sec 1.80 GBytes 15.5 Gbits/sec
[ 5] 1.00-2.00 sec 1.85 GBytes 15.9 Gbits/sec
[ 5] 2.00-3.00 sec 1.88 GBytes 16.2 Gbits/sec
[ 5] 3.00-4.00 sec 1.87 GBytes 16.1 Gbits/sec
[ 5] 4.00-5.00 sec 1.88 GBytes 16.2 Gbits/sec
[ 5] 5.00-6.00 sec 1.93 GBytes 16.6 Gbits/sec
[ 5] 6.00-7.00 sec 2.00 GBytes 17.2 Gbits/sec
[ 5] 7.00-8.00 sec 1.93 GBytes 16.6 Gbits/sec
[ 5] 8.00-9.00 sec 1.86 GBytes 16.0 Gbits/sec
[ 5] 9.00-10.00 sec 1.95 GBytes 16.8 Gbits/sec
[ 5] 10.00-10.04 sec 85.2 MBytes 17.3 Gbits/sec
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate
[ 5] (sender statistics not available)
[ 5] 0.00-10.04 sec 19.0 GBytes 16.3 Gbits/sec receiver
rcv_tcp_congestion cubic
iperf 3.9
Linux ps404a-3 5.15.137-pserver #5.15.137-6~deb11 SMP Thu Jan 4
07:19:34 UTC 2024 x86_64
-----------------------------------------------------------
Server listening on 5201
-----------------------------------------------------------
^Ciperf3: interrupt - the server has terminated
1 jwang@xxxxxxxxxxxx:~$ sudo ib_send_bw -F -a
************************************
* Waiting for client to connect... *
************************************
---------------------------------------------------------------------------------------
Send BW Test
Dual-port : OFF Device : mlx5_0
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
PCIe relax order: ON
ibv_wr* API : ON
RX depth : 512
CQ Moderation : 100
Mtu : 4096[B]
Link type : IB
Max inline data : 0[B]
rdma_cm QPs : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0x24 QPN 0x0174 PSN 0x300138
remote address: LID 0x17 QPN 0x004a PSN 0xc54d6f
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
2 1000 0.00 6.46 3.385977
4 1000 0.00 10.38 2.721894
8 1000 0.00 25.69 3.367830
16 1000 0.00 41.46 2.716859
32 1000 0.00 102.98 3.374577
64 1000 0.00 206.12 3.377053
128 1000 0.00 405.03 3.318007
256 1000 0.00 821.52 3.364939
512 1000 0.00 2150.78 4.404803
1024 1000 0.00 4288.13 4.391044
2048 1000 0.00 8518.25 4.361346
4096 1000 0.00 11440.77 2.928836
8192 1000 0.00 11526.45 1.475385
16384 1000 0.00 11526.06 0.737668
32768 1000 0.00 11524.86 0.368795
65536 1000 0.00 11331.84 0.181309
131072 1000 0.00 11524.75 0.092198
262144 1000 0.00 11525.82 0.046103
524288 1000 0.00 11524.70 0.023049
1048576 1000 0.00 11510.84 0.011511
2097152 1000 0.00 11524.58 0.005762
4194304 1000 0.00 11514.26 0.002879
8388608 1000 0.00 11511.01 0.001439
---------------------------------------------------------------------------------------
you can see with ipoib, it reaches 16 Gb/s using TCP, 1 streams,
131072 byte blocks
with RDMA at 4k+ message size it reaches 100 Gb/s
>
> >
> > Switching back to TCP will lead us to the old problems which was
> > solved by RDMA migration.
>
> Can you elaborate the problems, and why tcp won't work in this case? They
> may not be directly relevant to the issue we're discussing, but I'm happy
> to learn more.
>
> What is the NICs you were testing before? Did the test carry out with
> things like modern ones (50Gbps-200Gbps NICs), or the test was done when
> these hardwares are not common?
We use Mellanox/NVidia IB HCA from 40 Gb/s to 200 Gb/s mixed
generation across globe.
>
> Per my recent knowledge on the new Intel hardwares, at least the ones that
> support QPL, it's easy to achieve single core 50Gbps+.
In good case, I've also seen 50 Gbps + on Mellanox HCA.
>
> https://lore.kernel.org/r/PH7PR11MB5941A91AC1E514BCC32896A6A3342@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
>
> Quote from Yuan:
>
> Yes, I use iperf3 to check the bandwidth for one core, the bandwith is 60Gbps.
> [ ID] Interval Transfer Bitrate Retr Cwnd
> [ 5] 0.00-1.00 sec 7.00 GBytes 60.1 Gbits/sec 0 2.87 MBytes
> [ 5] 1.00-2.00 sec 7.05 GBytes 60.6 Gbits/sec 0 2.87 Mbytes
>
> And in the live migration test, a multifd thread's CPU utilization is almost 100%
>
> It boils down to what old problems were there with tcp first, though.
Yeah, this is the key reason we use RDMA. (low cpu ulitization and
better performance)
>
> >
> > >
> > > Per our best knowledge, RDMA users are rare, and please let anyone know if
> > > you are aware of such users. IIUC the major reason why RDMA stopped being
> > > the trend is because the network is not like ten years ago; I don't think I
> > > have good knowledge in RDMA at all nor network, but my understanding is
> > > it's pretty easy to fetch modern NIC to outperform RDMAs, then it may make
> > > little sense to maintain multiple protocols, considering RDMA migration
> > > code is so special so that it has the most custom code comparing to other
> > > protocols.
> > +cc some guys from Huawei.
> >
> > I'm surprised RDMA users are rare, I guess maybe many are just
> > working with different code base.
>
> Yes, please cc whoever might be interested (or surprised.. :) to know this,
> and let's be open to all possibilities.
>
> I don't think it makes sense if there're a lot of users of a feature then
> we deprecate that without a good reason. However there's always the
> resource limitation issue we're facing, so it could still have the
> possibility that this gets deprecated if nobody is working on our upstream
> branch. Say, if people use private branches anyway to support rdma without
> collaborating upstream, keeping such feature upstream then may not make
> much sense either, unless there's some way to collaborate. We'll see.
Is there document/link about the unittest/CI for migration tests, Why
are those tests missing?
Is it hard or very special to set up an environment for that? maybe we
can help in this regards.
>
> It seems there can still be people joining this discussion. I'll hold off
> a bit on merging this patch to provide enough window for anyone to chim in.
Thx for discussion and understanding.
Jinpu Wang
>
> Thanks,
>
> > >
> > > Thanks,
> > >
> > > --
> > > Peter Xu
> >
> > Thx!
> > Jinpu Wang
> > >
> >
>
> --
> Peter Xu
>
_______________________________________________
Devel mailing list -- devel@xxxxxxxxxxxxxxxxx
To unsubscribe send an email to devel-leave@xxxxxxxxxxxxxxxxx
