[Lustre-discuss] Could "Client & MGS" or "Client & OSS" running at one node together?

[Michael Shuey]

I've had success with LACP bonding, with an LACP-aware switch.  You
may also want to check your xmit_hash_policy (it's a kernel option to
the linux bonding driver).  I've had the best luck with layer3+4
bonding, using several OSSes with sequential IPs, and striping files
at a multiple of the number of links in the bond.

With most bonding modes, packets can get send across different links
in the bond.  This results in out-of-order packets, and can slow down
a TCP stream (like your Lustre connection) unnecessarily.  LACP will
route packets to a given destination across exactly one link in the
bond - but that will limit each TCP stream to the link speed of a
single bond member.

You can improve upon single link speeds with Lustre, because Lustre
will let you stripe large files across multiple OSSes.  A client will
build a separate TCP connection to each OSS, so as long as traffic
passes over different links you can use all the available bandwidth.
The way traffic is scattered across the bond members is controlled by
the xmit_hash_policy option; layer3+4 uses an XOR of the source and
destination addresses, combined with the source and destination TCP
ports (modulo number of links in the bond) to pick the specific link
for that stream.  If you're using sequential IPs for your OSSes, you
should be able to get a good scattering effect (since your source
address and port won't change, but your destination address will vary
across the OSSes).

A few years ago, I was using Lustre 1.4 and 1.6 and saw 200+ MB/sec
across two gigE links bonded together on the client (using striped
files).  Your mileage may vary, of course.  Caveats include:

Small files may have poorer performance than usual, due to transaction
overhead to multiple OSSes (if your bond is on the client).
Similarly, non-striped files will only see the speed of a single link.

If your OSTs start to fill, Lustre's load balancing may not give you
an ideal distribution of stripes across OSSes - causing multiple TCP
streams to land on the same bond member on the client.  Unfortunately,
this will present as slowdowns for certain files on certain clients
(because the number of bonds that can be used is a function of both
which OSSes are used in the file and the client's IP in the hash
policy).

All metadata accesses are limited to the speed of a single bond member
on the client.

If your bonds are on the server, then (as long as you have a number of
clients) you should see a nice increase in overall IO throughput.  It
won't be as marked a boost as 10gigE or Infiniband, but bonds are
inexpensive and generally better than a single link (to multiple
clients).

Hope this helps - good luck!

--
Mike Shuey



On Wed, Mar 7, 2012 at 9:27 PM, zhengfeng <zf5984599 at gmail.com> wrote:
> Dear all, thanks a lot for your answers ;)
>
> Now I have another problem about the network between nodes.
> Since there is no Infiniband or 10G-NIC, but I still want to
> increase the bandwidth by add more 1G-NICs, I plan to use Linux bonding.
>
> Then, bonding 4 NICs together at one node, BUT there is NO performance
> enhanced no matter which bongding mode, described in kernel doc, used.
> In stead, the performance of 4-NICs-bonding is lower than 1 NIC's.
> Then we use 2 NICs bonding, the performance is better than 1 NIC's.
> The result is :
> bonding 2 NIC: 1 + 1 > 1
> bonding 4 NIC: 1 + 1 + 1 + 1 < 1
>
> So confused......
> The benchmark we used is "netperf".
> And I use "tcpdump" to dump the packages, found that there are great
> of TCP segments out of orders.
>
> My question is that:
> a) TCP segments are out of order, which induced that 4-NIC-bonding
>  performance decay, is this the root cause?
> b) We are doubting the feasibility of this method: using 4-NIC-bonding
> to increase bandwidth. Any proposals about that?  If so, maybe I should
> use some other method instead of this.
>
> Thanks again, all
>
> ________________________________
> Best Regards
> Zheng
>
> From: Peter Grandi
> Date: 2012-03-07 20:54
> To: Lustre discussion
> Subject: Re: [Lustre-discuss] Could "Client & MGS" or "Client & OSS" running
> at one node together?
>> Since there is no more node in our project when using Lustre,
>> I want to confirm that:
>
>> 1) Could the "Client" and "MGS" run at one node together? or
>> could "Client" and "OSS" run at one node together? 2) Suppose
>> I had deployed them at one node, what potential shortcomings
>> or harm are there?
>
> Running MGS and MDS on the same nodes is customary, see:
>    http://wiki.lustre.org/manual/LustreManual20_HTML/LustreOperations.html#50438194_24122
>
> Running the MGS, MDS and OSS service on the same node is
> possible and fairly common in very small setups, usually those
> in which there is only 1-2 nodes.
>
> It is possible to use the client code on all types of Lustre
> servers, but at least in the case of using the client code on an
> OSS there is the non-negligible possibility of a resource
> deadlock, if the client uses the OSS on the same node, as the
> client and OSS codes compete for memory, so in the past this has
> been discouraged.
>
> This is documented here:
>   http://wiki.lustre.org/manual/LustreManual20_HTML/LustreOperations.html#50438194_84876
>    «Caution - Do not do this when the client and OSS are on the
>     same node, as memory pressure between the client and OSS can
>     lead to deadlocks.»
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-discuss
>
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-discuss
>