On Mon, 26 Jul 2010, Greg Smith wrote:
> Matthew Wakeling wrote:
>> Does your latency graph really have milliseconds as the y axis? If so, this
>> device is really slow - some requests have a latency of more than a second!
>
> Have you tried that yourself? If you generate one of those with standard
> hard drives and a BBWC under Linux, I expect you'll discover those latencies
> to be >5 seconds long. I recently saw >100 *seconds* running a large pgbench
> test due to latency flushing things to disk, on a system with 72GB of RAM.
> Takes a long time to flush >3GB of random I/O out to disk when the kernel
> will happily cache that many writes until checkpoint time.
Apologies, I was interpreting the graph as the latency of the device, not
all the layers in-between as well. There isn't any indication in the email
with the graph as to what the test conditions or software are. Obviously
if you factor in checkpoints and the OS writing out everything, then you
would have to expect some large latency operations. However, if the device
itself behaved as in the graph, I would be most unhappy and send it back.
Yeb also made the point - there are far too many points on that graph to
really tell what the average latency is. It'd be instructive to have a few
figures, like "only x% of requests took longer than y".
Matthew
--
I wouldn't be so paranoid if you weren't all out to get me!!
