Hi, I'm working on the QCDOC massively parallel supercomputer project. It's a custom designed 10Tflop MPP backend based on a custom PPC440 based SoC, with a custom internode serial communications network, and plain old ethernet boot/diagnostics/IO auxiliary tree. (see http://www.physics.columbia.edu/~cqft/) We're round robin sending to > 1024 IP's/MAC's on a subnet, kernel version (2.4.20-31.9). I'm observing UDP "sendto" returning ENOBUFS (the man page says this is impossible). No arp's go out for the failing IP address, and reading the kernel code suggests ENOBUFS can be returned when neigh_alloc fails. Is there any way to prevent this? Subsequent re-transmit attempts to this IP continue to return ENOBUFS. It can take many many seconds until these to succeed. (is this the GC timeout?). It seems to me that the old arp entries should be evicted and replaced rather than having neigh_alloc fail? (Of course, I guess I can decrease the GC interval, but I would have though active eviction (or even GC invocation) neigh_alloc failure would makes sense)? Is my interpretation correct? Many thanks, Peter - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
