Hello,
Although the SRP protocol supports multichannel operation, although
since considerable time RDMA HCA's are available that support multiple
completion vectors and although multichannel operation yields better
performance than using a single channel, the Linux SRP initiator does
not yet support multichannel operation. While adding multichannel
support in the SRP initiator I encountered a few challenges of which I
think these need wider discussion. The topics I would like invite wider
discussion about are as follows:
- How to avoid unneeded inter-socket cache traffic. Should the blk-mq
layer e.g. assign CPU cores to hardware queues such that all CPU cores
associated with a single hardware queue reside on the same CPU socket?
(see also patch 1/8)
- How to pass the hardware context selected by the block layer to the
SCSI LLD queuecommand() callback function ? (see also patches 2/8 and
3/8).
- Which approach should a SCSI LLD follow for selection of an MSI-X
completion vector to ensure that the interrupt handler is invoked on
the same CPU socket as the blk-mq hardware context data structures ?
As one can see patch 8/8 relies on the assumption that completion
vectors have been spread evenly over CPU sockets. If a HCA e.g.
supports eight completion vectors then that means that in a system
with two CPU sockets vectors 0-3 are associated with a CPU core on
the first CPU socket and vectors 4-7 with a CPU core on the second CPU
socket.
The patches in this series are:
0001-blk-mq-Use-all-available-hardware-queues.patch
0002-scsi-mq-Add-support-for-multiple-hardware-queues.patch
0003-scsi-mq-Pass-hctx-to-low-level-SCSI-drivers.patch
0004-IB-srp-Move-ib_destroy_cm_id-call-into-srp_free_ch_i.patch
0005-IB-srp-Remove-stale-connection-retry-mechanism.patch
0006-IB-srp-Avoid-that-I-O-hangs-due-to-a-cable-pull-duri.patch
0007-IB-srp-Separate-target-and-channel-variables.patch
0008-IB-srp-Add-multichannel-support.patch
Note: a predecessor of this patch series has been used to measure the
performance of Christoph's scsi-mq patches that have been merged in
kernel version v3.17-rc1.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html