Re: Re: reduce iSERT Max IO size

[Krishnamraju Eraparaju <krishna2@chelsio.com>]

On Thursday, October 10/08/20, 2020 at 13:12:39 +0000, Bernard Metzler wrote:
> -----"Krishnamraju Eraparaju" <krishna2@chelsio.com> wrote: -----
> 
> >To: "Max Gurtovoy" <mgurtovoy@nvidia.com>
> >From: "Krishnamraju Eraparaju" <krishna2@chelsio.com>
> >Date: 10/07/2020 05:36AM
> >Cc: "Sagi Grimberg" <sagi@grimberg.me>, linux-rdma@vger.kernel.org,
> >"Potnuri Bharat Teja" <bharat@chelsio.com>, "Max Gurtovoy"
> ><maxg@mellanox.com>
> >Subject: [EXTERNAL] Re: reduce iSERT Max IO size
> >
> >On Sunday, October 10/04/20, 2020 at 00:45:26 +0300, Max Gurtovoy
> >wrote:
> >> 
> >> On 10/3/2020 6:36 AM, Krishnamraju Eraparaju wrote:
> >> >On Friday, October 10/02/20, 2020 at 13:29:30 -0700, Sagi Grimberg
> >wrote:
> >> >>>Hi Sagi & Max,
> >> >>>
> >> >>>Any update on this?
> >> >>>Please change the max IO size to 1MiB(256 pages).
> >> >>I think that the reason why this was changed to handle the worst
> >case
> >> >>was in case there are different capabilities on the initiator and
> >the
> >> >>target with respect to number of pages per MR. There is no
> >handshake
> >> >>that aligns expectations.
> >> >But, the max pages per MR supported by most adapters is around 256
> >pages
> >> >only.
> >> >And I think only those iSER initiators, whose max pages per MR is
> >4096,
> >> >could send 16MiB sized IOs, am I correct?
> >> 
> >> If the initiator can send 16MiB, we must make sure the target is
> >> capable to receive it.
> >I think max IO size, at iSER initiator, depends on
> >"max_fast_reg_page_list_len".
> >currently, below are the supported "max_fast_reg_page_list_len" of
> >various iwarp drivers:
> >
> >iw_cxgb4: 128 pages
> >Softiwarp: 256 pages
> >i40iw: 512 pages
> >qedr: couldn't find.
> >
> 
> For siw, this limit is not determined by resource constraints.
> We could bump it up to 512, or higher. What is a reasonable
> maximum, from iSER view?

If the most common IO sizes are 4K & 8K, then the reasonable max IO size of
256 pages(1 MiB) would be appropriate, by default. currently, NVMet-rdma
also limits max IO size to 1MiB.
> 
> 
> >For iwarp case, if 512 is the max pages supported by all iwarp
> >drivers,
> >then provisioning a gigantic MR pool at target(to accommodate never
> >used
> >16MiB IO) wouldn't be a overkill?
> >> 
> >> >
> >> >>If we revert that it would restore the issue that you reported in
> >the
> >> >>first place:
> >> >>
> >> >>--
> >> >>IB/isert: allocate RW ctxs according to max IO size
> >> >I don't see the reported issue after reducing the IO size to 256
> >> >pages(keeping all other changes of this patch intact).
> >> >That is, "attr.cap.max_rdma_ctxs" is now getting filled properly
> >with
> >> >"rdma_rw_mr_factor()" related changes, I think.
> >> >
> >> >Before this change "attr.cap.max_rdma_ctxs" was hardcoded with
> >> >128(ISCSI_DEF_XMIT_CMDS_MAX) pages, which is very low for single
> >target
> >> >and muli-luns case.
> >> >
> >> >So reverting only ISCSI_ISER_MAX_SG_TABLESIZE macro to 256 doesn't
> >cause the
> >> >reported issue.
> >> >
> >> >Thanks,
> >> >Krishnam Raju.
> >> >>Current iSER target code allocates MR pool budget based on queue
> >size.
> >> >>Since there is no handshake between iSER initiator and target on
> >max IO
> >> >>size, we'll set the iSER target to support upto 16MiB IO
> >operations and
> >> >>allocate the correct number of RDMA ctxs according to the factor
> >of MR's
> >> >>per IO operation. This would guaranty sufficient size of the MR
> >pool for
> >> >>the required IO queue depth and IO size.
> >> >>
> >> >>Reported-by: Krishnamraju Eraparaju <krishna2@chelsio.com>
> >> >>Tested-by: Krishnamraju Eraparaju <krishna2@chelsio.com>
> >> >>Signed-off-by: Max Gurtovoy <maxg@mellanox.com>
> >> >>--
> >> >>
> >> >>>
> >> >>>Thanks,
> >> >>>Krishnam Raju.
> >> >>>On Wednesday, September 09/23/20, 2020 at 01:57:47 -0700, Sagi
> >Grimberg wrote:
> >> >>>>>Hi,
> >> >>>>>
> >> >>>>>Please reduce the Max IO size to 1MiB(256 pages), at iSER
> >Target.
> >> >>>>>The PBL memory consumption has increased significantly after
> >increasing
> >> >>>>>the Max IO size to 16MiB(with commit:317000b926b07c).
> >> >>>>>Due to the large MR pool, the max no.of iSER connections(On
> >one variant
> >> >>>>>of Chelsio cards) came down to 9, before it was 250.
> >> >>>>>NVMe-RDMA target also uses 1MiB max IO size.
> >> >>>>Max, remind me what was the point to support 16M? Did this
> >resolve
> >> >>>>an issue?
> >
>