Re: Latency of "direct path reads"

From: Austin Hackett <hacketta_57_at_me.com>
Date: Wed, 21 Aug 2013 20:25:36 +0100
Message-id: <4D50ABBF-0B28-4ACA-A0D2-AD41984C18DB_at_me.com>



Hi Amir

You're welcome.

Regarding my experience so far with dNFS:

Fortunately, I was able to spend some time testing direct NFS in a lab environment using SLOB and other tools. In addition to the bug I mentioned before, I also hit 14128555. There was also a "problem" with multi path failover, but it was really to a mis-understanding. From my reading of the manuals etc. I'd concluded that I could configure two paths with 2 un-bonded NICs and failover would be handled by dNFS. Whilst both paths were used, failover didn't work in both directions. I hadn't realised that the KNFS mount used by DNFS has to be OS bonded in order for failover to work properly. However, it took various support escalations and push back on inappropriate advice before I was able to determine the details of how multi path should actually be deployed. That experience left me with the feeling it's still pretty niche and not well understood within Oracle Support.

During the lab tests, I was able to perform the same tests against the same storage, configured with Fibre Channel and ASM instead of NFS/dNFS. Whilst FC did have a slight edge in terms of the IOPS I could drive, dNFS could still deliver way more IOPS than our app would ever need with low enough latency. I seem to recall FC/ASM managed around 20,000 IOPS more, but dNFS still delivered into the 6-digits.

Now we're in production, and I have to say that I'm pretty happy with dNFS.

On 20 Aug 2013, at 03:14, "Hameed, Amir" <Amir.Hameed_at_xerox.com> wrote:

> Hi Austin,
> I really appreciate your feedback. I did go through the document that you have mentioned below a few times in the past but most of the patches listed at that time were either released for Linux or Exadata for Solaris; we are SPARC Solaris. After we provided pstack dumps to Oracle today, Oracle confirmed that we are hitting the same bug that you have mentioned below. But since we are at 11.2.0.3.6 and the patch for Solaris is available for 11.2.0.3.5, Oracle is in the process of back-porting the patch.
>
> You are also right about the 32k chunk size because of the rsize and wsize mount options and I confirmed it by running truss. How has your experience been with dNFS so far?
>
> Thanks

--
http://www.freelists.org/webpage/oracle-l
Received on Wed Aug 21 2013 - 21:25:36 CEST

Original text of this message