Roland> My plan is to change the receive handling of IPoIB
Roland> slightly, so that if it can't allocate a new receive
Roland> buffer, it reposts the old buffer and drops the packet it
Roland> just received.
Here's a patch that changes IPoIB to use this scheme. This should be
much more robust when the system gets low on GFP_ATOMIC memory.
I'd appreciate it if people could stress test and benchmark this. It
works well for me, but I'm wondering if this patch has any effect on
performance (either better or worse).
Helen, it would be especially interesting if you could run your test
with this patch and without increasing min_free_kbytes, since you are
able to reproduce GFP_ATOMIC failures. I'd be curious to know what
you see in /sys/class/net/ib0/statistics/rx_dropped after running the test.
As a general rule, dropping a packet that has traversed the network is frowned upon by the IETF (this is perhaps more due to the view of the network being all of IP and not just within a data center). I understand the idea but given it is UD, the HCA can effectively drop the packet without causing any side effects which should result in lower host CPU / I/O / memory utilitization.
Mike
_______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
