[gpfsug-discuss] Spectrum Scale Ganesha NFS multi threaded AFM?
Andi Christiansen
andi at christiansen.xxx
Mon Feb 24 22:31:45 GMT 2020
Hi all,
Thank you for all your suggestions! The latency is 30ms between the sites (1600km to be exact). So if I have entered correctly in the calculator 1Gb is actually what is expected on that distance.
I had a meeting today with IBM where we were able to push that from the 1Gb to about 4Gb on one link with minimal tuning, more tuning will come the next few days! We are also looking to implement the feature afmParallelMounts which should give us the full bandwidth we have between the sites :-)
Thanks!
Best Regards
Andi Christiansen
Sendt fra min iPhone
> Den 22. feb. 2020 kl. 10.35 skrev Tomer Perry <TOMP at il.ibm.com>:
>
> Hi,
>
> Its implied in the tcp tuning suggestions ( as one needs bandwidth and latency in order to calculate the BDP).
> The overall theory is documented in multiple places (tcp window, congestion control etc.) - nice place to start is https://en.wikipedia.org/wiki/TCP_tuning.
> I tend to use this calculator in order to find out the right values https://www.switch.ch/network/tools/tcp_throughput/
>
> The parallel IO and multiple mounts are on top of the above - not instead ( even though it could be seen that it makes things better - but multiple of the small numbers we're getting initially).
>
> Regards,
>
> Tomer Perry
> Scalable I/O Development (Spectrum Scale)
> email: tomp at il.ibm.com
> 1 Azrieli Center, Tel Aviv 67021, Israel
> Global Tel: +1 720 3422758
> Israel Tel: +972 3 9188625
> Mobile: +972 52 2554625
>
>
>
>
> From: "Luis Bolinches" <luis.bolinches at fi.ibm.com>
> To: "gpfsug main discussion list" <gpfsug-discuss at spectrumscale.org>
> Cc: Jake Carrol <jake.carroll at uq.edu.au>
> Date: 22/02/2020 07:56
> Subject: [EXTERNAL] Re: [gpfsug-discuss] Spectrum Scale Ganesha NFS multi threaded AFM?
> Sent by: gpfsug-discuss-bounces at spectrumscale.org
>
>
>
> Hi
>
> While I agree with what es already mention here and it is really spot on, I think Andi missed to reveal what is the latency between sites. Latency is as key if not more than ur pipe link speed to throughput results.
>
> --
> Cheers
>
> On 22. Feb 2020, at 3.08, Andrew Beattie <abeattie at au1.ibm.com> wrote:
>
> Andi,
>
> You may want to reach out to Jake Carrol at the University of Queensland,
>
> When UQ first started exploring with AFM, and global AFM transfers they did extensive testing around tuning for the NFS stack.
>
> From memory they got to a point where they could pretty much saturate a 10GBit link, but they had to do a lot of tuning to get there.
>
> We are now effectively repeating the process, with AFM but using 100GB links, which brings about its own sets of interesting challenges.
>
>
>
>
>
> Regards
>
> Andrew
>
> Sent from my iPhone
>
> On 22 Feb 2020, at 09:32, Andi Christiansen <andi at christiansen.xxx> wrote:
>
> Hi,
>
> Thanks for answering!
>
> Yes possible, I’m not too much into NFS and AFM so I might have used the wrong term..
>
> I looked at what you suggested (very interesting reading) and setup multiple cache gateways to our home nfs server with the new afmParallelMount feature. It was as I suspected, for each gateway that does a write it gets 50-60MB/s bandwidth so although this utilizes more when adding it up (4 x gateways = 4 x 50-60MB/s) I’m still confused to why one server with one link cannot utilize more than the 50-60MB/s on 10Gb links ? Even 200-240MB/s is much slower than a regular 10Gbit interface.
>
> Best Regards
> Andi Christiansen
>
>
>
> Sendt fra min iPhone
>
> Den 21. feb. 2020 kl. 18.25 skrev Tomer Perry <TOMP at il.ibm.com>:
>
> Hi,
>
> I believe the right term is not multithreaded, but rather multistream. NFS will submit multiple requests in parallel, but without using large enough window you won't be able to get much of each stream.
> So, the first place to look is here: https://www.ibm.com/support/knowledgecenter/en/STXKQY_5.0.4/com.ibm.spectrum.scale.v5r04.doc/bl1adm_tuningbothnfsclientnfsserver.htm- and while its talking about "Kernel NFS" the same apply to any TCP socket based communication ( including Ganesha). I tend to test the performance using iperf/nsdperf ( just make sure to use single stream) in order to see what is the expected maximum performance.
> After that, you can start looking into "how can I get multiple streams?" - for that there are two options:
> https://www.ibm.com/support/knowledgecenter/en/STXKQY_5.0.4/com.ibm.spectrum.scale.v5r04.doc/bl1ins_paralleldatatransfersafm.htm
> and
> https://www.ibm.com/support/knowledgecenter/en/STXKQY_5.0.4/com.ibm.spectrum.scale.v5r04.doc/b1lins_afmparalleldatatransferwithremotemounts.htm
>
> The former enhance large file transfer, while the latter ( new in 5.0.4) will help with multiple small files as well.
>
>
>
> Regards,
>
> Tomer Perry
> Scalable I/O Development (Spectrum Scale)
> email: tomp at il.ibm.com
> 1 Azrieli Center, Tel Aviv 67021, Israel
> Global Tel: +1 720 3422758
> Israel Tel: +972 3 9188625
> Mobile: +972 52 2554625
>
>
>
>
> From: Andi Christiansen <andi at christiansen.xxx>
> To: "gpfsug-discuss at spectrumscale.org" <gpfsug-discuss at spectrumscale.org>
> Date: 21/02/2020 15:25
> Subject: [EXTERNAL] [gpfsug-discuss] Spectrum Scale Ganesha NFS multi threaded AFM?
> Sent by: gpfsug-discuss-bounces at spectrumscale.org
>
>
>
> Hi all,
>
> i have searched the internet for a good time now with no answer to this.. So i hope someone can tell me if this is possible or not.
>
> We use NFS from our Cluster1 to a AFM enabled fileset on Cluster2. That is working as intended. But when AFM transfers files from one site to another it caps out at about 5-700Mbit/s which isnt impressive.. The sites are connected on 10Gbit links but the distance/round-trip is too far/high to use the NSD protocol with AFM.
>
> On the cluster where the fileset is exported we can only see 1 session against the client cluster, is there a way to either tune Ganesha or AFM to use more threads/sessions?
>
> We have about 7.7Gbit bandwidth between the sites from the 10Gbit links and with multiple NFS sessions we can reach the maximum bandwidth(each using about 50-60MBits per session).
>
> Best Regards
> Andi Christiansen _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
>
>
>
>
>
> Ellei edellä ole toisin mainittu: / Unless stated otherwise above:
> Oy IBM Finland Ab
> PL 265, 00101 Helsinki, Finland
> Business ID, Y-tunnus: 0195876-3
> Registered in Finland
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
>
>
>
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20200224/0033fa75/attachment-0002.htm>
More information about the gpfsug-discuss
mailing list