[Beowulf] NFS over RDMA performance confusion

[holway at th.physik.uni-frankfurt.de]

>
> Ah, this looks great!  You added 50% IOPs with the doubling of procs,
> and I would bet you could squeeze a little more out by going to 24 or 32
> procs.

	Children see throughput for 1 random writers 	=   44703.30 KB/sec
	Children see throughput for 2 random writers 	=   80244.89 KB/sec
	Children see throughput for 4 random writers 	=  123348.99 KB/sec
	Children see throughput for 8 random writers 	=  202080.27 KB/sec
	Children see throughput for 16 random writers =  287422.56 KB/sec
	Children see throughput for 24 random writers =  322134.46 KB/sec
	Children see throughput for 32 random writers =  307318.69 KB/sec
	Children see throughput for 64 random writers =  329277.73 KB/sec

	Children see throughput for 1 random readers =  115058.72 KB/sec
	Children see throughput for 2 random readers =  197125.98 KB/sec
	Children see throughput for 4 random readers =  373389.84 KB/sec
	Children see throughput for 8 random readers =  701358.48 KB/sec
	Children see throughput for 16 random readers = 1106813.94 KB/sec
	Children see throughput for 24 random readers = 1323350.00 KB/sec
	Children see throughput for 32 random readers = 1326093.64 KB/sec
	Children see throughput for 64 random readers = 1331465.57 KB/sec


Somewhat.. I have 16 physical cores on this machine btw.

These numbers seem awfully low however. Each process seems to be limited
to 7000 or so IOPS.



I was trying to test this in the local tempfs but I got an error....

Can not open temp file: iozone.tmp
open: Invalid argument

Not sure what is going on there.