Re: Orion... should we believe what we see?

From: Philip Jones <>
Date: Mon, 25 Feb 2013 00:50:39 +0000
Message-Id: <>

You really should use SLOB instead.


Kevin's blog should be in your list of RSS feeds if it isn't already!



On 25 Feb 2013, at 00:34, "Tornblad, John" <> wrote:

> We are seeing wildy varying performance in an Orion run meant to simulate a DW workload (large sequential 1MB reads).
> The output below seems to be "red flags" all over:
> a) Wildly varying results, even counterintuitive results (more concurrent reads = less throughput)
> b) Latency on the first test (1 concurrent read at a time) showed 34 IOs at a 1-2 sec latency
> c) In other tests, smaller reads (128KB) seem to perform better but have only yielded a maximum of ~1300 MB/sec
> Trying to employ some "USE" (utilization/saturation/errors) methodology but this is time consuming. There is some skepticism of Orion's reliability in our shop. Problem statement: we believe this frame should be producing closer to 2300 MB/sec bandwidth on large sequential IOs.
> Any comments regarding weirdness (or normality?) of these results, steps to take next, observations would be greatly appreciated.
> Orion's documentation is a little scant and not many (any?) metalink notes regarding effectively using it. I've listened to Alex Gorbachev talk about Orion a couple of times but need more input on what to do or think when "things don't go right".
> -

Received on Mon Feb 25 2013 - 01:50:39 CET

Original text of this message