On 16:30 Mon 21 Oct, Doug Ledford wrote: > On Mon, 2019-09-16 at 17:42 +0800, Liu, Changcheng wrote: > > Hi all, > > I'm working on using rdma to improve message transaction > qperf is nice because it will do both the tcp and rdma testing, so the > same set of options will make it behave the same way under both tests. @Doug Ledford: I'll check how to use it to compare RDMA & TCP. > > I think you are mis-reading the instructions on ib_send_bw. First of > all, IB RC queue pairs are, when used in send/recv mode, message passing > devices, not a stream device. When you specified the -s parameter of @Doug Ledford: What's the difference between "message passing device" and "stream device"? > 1GB, you were telling it to use messages of 1GB in size, not to pass a > total of 1GB of messages. And the default number of messages to pass is > 1,000 iterations (the -n or --iters options), so you were actually @Doug Ledford: Thanks for your information. It helps me a lot. > testing a 1,000GB transfer. You would be better off to use a smaller > message size and then set the iters to the proper value. This is what I > got with 1000 iters and 1GB message size: > > #bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps] > 1073741824 1000 6159.64 6159.46 0.000006 > --------------------------------------------------------------------------------------- > > real 3m3.101s > user 3m2.430s > sys 0m0.450s > > I tried dropping it to 1 iteration to make a comparison, but that's not > even allowed by ib_send_bw, it wants a minimum of 5 iterations. So I > did 8 iterations at 1/8th GB in size and this is what I got: > > #bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps] > 134217728 8 6157.54 6157.54 0.000048 > --------------------------------------------------------------------------------------- > > real 0m2.506s > user 0m2.411s > sys 0m0.059s > > When I adjust that down to 1MB and 1024 iters, I get: > > #bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps] > 1048576 1024 6157.74 6157.74 0.006158 > --------------------------------------------------------------------------------------- > > real 0m0.427s > user 0m0.408s > sys 0m0.002s > > The large difference in time between these last two tests, given that > the measured bandwidth is so close to identical, explains the problem > you are seeing below. > > The ib_send_bw test is a simple test. It sets up a buffer by > registering its memory, then just slams that buffer over the wire. With > a 128MB buffer, you pay a heavily memory registration penalty. That's > factored into the 2s time difference between the two runs. When you use > a 1GB buffer, the delay is noticeable to the human eye. There is a very > visible pause as the server and client start their memory registrations. @Doug Ledford: Do you mean that every RDMA-SGE(Scatter/Gather element) will use seperate MR(Memory Region)? If all the RDMA-SGE use only one pre-allocated MR-1GB, the two tests shouldn't have so much time consuming difference. > > > In Ceph, the result shows that rdma performance (RC transaction > > type, > > SEDN operation) is worse or not much better than TCP implemented > > performance. > > Test A: > > 1 client thread send 20GB data to 1 server thread (marked as > > 1C:1S) > > Result: > > 1) implementation based on RDMA > > Take 171.921294s to finish send 20GB data. > > 2) implementation based on TCP > > Take 62.444163s to finish send 20GB data. > > > > Test B: > > 16 client threads send 16x20GB data to 1 server thread (marked > > as 16C:1S) > > Result: > > 1) implementation base on RDMA > > Take 261.285612s to finish send 16x20GB data. > > 2) implementation based on TCP > > Take 318.949126 to finish send 16x20GB data. > > I suspect your performance problems here are memory registrations. As > noted by Chuck Lever in some of his recent postings, memory > registrations can end up killing performance for small messages, and the > tests I've shown here indicate, they're also a killer for huge memory > blocks if they are repeatedly registered/deregistered. TCP has no I think we could pre-registered 1GB MR and then all the SGE share with the same MR, then it could mitigate the penalty in register/deregister. > memory registration overhead, so in the single client case, it is > outperforming the RDMA case. But in the parallel case with lots of > clients, the memory registration overhead is spread out among many > clients, so we are able to perform better overall. In Ceph implementation, all the threads in the same process share with the same pre-registered 1GB MR. The MR is divided into lots of chunks to be used as SGE. In this way, how to explain the test result between Test-A & Test-B? > > In a nutshell, it sounds like the Ceph transfer engine over RDMA is not > optimized at all, and is hitting problems with memory registration > overhead. Ceph/RDMA seems not widely used and some implementation need to be optimized. I'm going to work on it in future. > > -- > Doug Ledford <dledford@xxxxxxxxxx> > GPG KeyID: B826A3330E572FDD > Fingerprint = AE6B 1BDA 122B 23B4 265B 1274 B826 A333 0E57 2FDD
