Shirley> Michael, Different workloads could use different
Shirley> send/recv queue sizes, also different HCAs could pick up
Shirley> different send/recv queue sizes to utilize the HCA
Shirley> performance. These parameters should be tunable instead
Shirley> of setting default values in IPoIB. Also Kevin pointed
Shirley> out, we don't want to waste the kernel memory when not
Shirley> necessary.
Hmm, it seems having a big send queue would be pretty cheap. Maybe we
should just increase the TX queue to something like 256 and be done
with it -- would you ever want a huge work queue? It seems that
setting the network stack's txqueuelen should be sufficient beyond
that length.
RX is a little trickier, because the receive buffers consume quite a
bit of memory. Maybe we should just create a big receive work queue
but not fill it by default? That would allow the RX queue to be
increased/decreased dynamically while the device is running.
By leaving the work queue sizes compile-time constants, we make the
code a tiny fraction more efficient.
- R.
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
