On Wed, Jul 03, 2019 at 11:56:04AM -0700, Sagi Grimberg wrote: > > > Hi Sagi, > > > > I'm not sharing your worries about bad out-of-the-box experience for a > > number of reasons. > > > > First of all, this code is part of upstream kernel and will take time > > till users actually start to use it as is and not as part of some distro > > backports or MOFED packages. > > True, but I am still saying that this feature is damaging sync IO which > represents the majority of the users. It might not be an extreme impact > but it is still a degradation (from a very limited testing I did this > morning I'm seeing a consistent 5%-10% latency increase for low QD > workloads which is consistent with what Yamin reported AFAIR). > > But having said that, the call is for you guys to make as this is a > Mellanox device. I absolutely think that this is useful (as I said > before), I just don't think its necessarily a good idea to opt it by > default given that only a limited set of users would take full advantage > of it while the rest would see a negative impact (even if its 10%). > > I don't have a hard objection here, just wanted to give you my > opinion on this because mlx5 is an important driver for rdma > users. Your opinion is very valuable for us and we started internal thread to challenge this "enable by default", it just takes time and I prefer to enable this code to get test coverage as wide as possible. > > > Second, Yamin did extensive testing and worked very close with Or G. > > and I have very high confident in the results of their team work. > > Has anyone tested other RDMA ulps? NFS/RDMA or SRP/iSER? > > Would be interesting to understand how other subsystems with different > characteristics behave with this. Me too, and I'll revert this default if needed. Thanks
