On Tue, May 17, 2011 at 8:43 PM, Venkateswararao Jujjuri (JV) <jv...@linux.vnet.ibm.com> wrote:
I sent comments on individual patches. This is turning out nice, many of the PDU handlers are now straight-line code that can be followed easily. > o Redesigned to use bh as per Stefan's suggestion. > This made the code very simple but is little less performant compared to V1. > Anthony suggested to go-in with cleaner code and design (This version) and > deal with the performance later. > Just to put in perspective: > Sequential Writes of creating 1GB files using ffsb > o Write size 8k > With bh: 66.9MB/sec > Without bh (marshalling routines): 74.9 MB/sec > > o Write size 128k > With bh: 117MB/sec > Without bh (marshalling routines): 122MB/sec It would be interesting to hack in a post-yield callback which is invoked after yield transfers control back to the calling coroutine. This avoids setting up a BH and notifying the iothread to invoke it. I'm just suggesting this as an experiment to see if the BH mechanism causes the performance degradation, not as a real solution. Stefan
