For some use case that there is heavy large read I/O, the performance
bottleneck due to MAX_LARGE_DATAIN_PER_CONNECTION was reported.
The following assumes that all I/Os are large read.
Large read primary task whose I/O size is more than
SPDK_BDEV_LARGE_BUF_MAX_SIZE (=64KB) is split into multiple
read subtasks.
spdk_iscsi_globals::MaxQueueDepth limits maximum number of outstanding
read primary tasks, and MAX_LARGE_DATAIN_PER_CONNECTION (=64)
limits maximum number of outstanding read subtasks.
MAX_LARGE_DATAIN_PER_CONNECTION is also used to calculate PDU pool.
To remove the performance bottleneck, change the macro constant
MAX_LARGE_DATAIN_PER_CONNECTION to a global variable
spdk_iscsi_globals::MaxLargeDataInPerConnection.
We don't see any negative side effect if we set
spdk_iscsi_globals::MaxLargeDataInPerConnection to 64.
The use case that reported the performance issue will change the
value of spdk_iscsi_globals::MaxLargeDataInPerConnection by its own
responsibility.
The next patch will add the value of
spdk_iscsi_globals::MaxLargeDataInPerConnection to iSCSI options,
and make it configurable by JSON RPC.
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Change-Id: Ifc30cdb8e00d50f4d3755ff399263cf5d0b681b6
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/3755
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
5b2c76f062