| Hi Rupert,
If you try RMDA (or work on the older, now removed RMDA code in glusterfs that use(d) ibverbs, be aware that there was a glitch, at least as of glusterfs v3.12.7: https://lists.gluster.org/pipermail/gluster-users/2018-May/034135.html A directory with tens of thousands of files did not show all directory entries.) I couldn't fix this back then and instead switched to ipoib. Although the measured per-file bandwith back then was better with RDMA than with ipoib by about 50%, I stayed with ipoib because in the aggreate bandwith (many clients reading many files simultaneously) it didn't make a difference. (Our bottlenecks are elsewhere, not in the network bandwidth between server and client.) I've used ipoib since then to run a glusterfs with more than 2 PB and haven't had serious problems so far.
best wishes, Stefan
Hi Rupert, Apologies for the inconvenience, even though this was discussed, somehow it got missed in official release notes or documentation.
Few old discussions are below:
Will surely document this as part of Gluster docs on an appropriate heading.
As evident from the above discussions, we deprecated it since we had no means to support it. If that changes, we'll be happy to work on re-enabling it, maybe as a compile-time option.
Thanks Saju
What surprised me is more that, if not looking into the developer channels, there is no indication that the feature is gone till it crashes when you try to use it. I did not think of
Gluster as a developer-only tool, even in the “community version”. Native RDMA support was the main reason for me to choose GlusterFS over MooseFS – although it’s just a feature on “paper” and I do not really know yet the real needs and bottlenecks
of our new cluster. With both RDMA and tiering gone, I based my decision on completely outdated information about gluster features. @strahil: I hope to find some time to try your suggestion next week, thanks. But even if it works, I will most probably not risk using something the developers consider not fit for use
anymore. There seem also to be issues with mixed mode and multiple IP addresses anyway – our computation and storage nodes are on a infiniband (RDMA and IP over Infiniband) island network, but all are also connected to the outside via IP over Ethernet. The
login nodes, which would also need access to the cluster file system are in a VM and only accessible via IP over Ethernet Rupert
As everything deprecated, it never get's completely removed from source.
Can you try to 'make' from source and check if v9 by default has 'Infiniband verbs' disabled ?
If yes, then you can compile with enabled supportand try that.
Strahil Nikolov
On Tue, Feb 23, 2021 at 10:40, Schlick Rupert
________ Community Meeting Calendar: Schedule - Every 2nd and 4th Tuesday at 14:30 IST / 09:00 UTC Bridge: https://meet.google.com/cpu-eiue-hvkGluster-users mailing list Gluster-users@xxxxxxxxxxxhttps://lists.gluster.org/mailman/listinfo/gluster-users
|
