Welcome first of all, glad to see interest in improving Gluster performance.
I believe there's a huge potential for improvements in different areas you can look into.
Certainly zero-copy (or just eliminating extra copies we have today in the code) would be most welcome.
We seem to have some lock contention in some areas of Gluster. Depending on the setup and workload, there might be different ones.
Certainly reducing some of them could lead to reduced latency and CPU consumption.
RDMA has been dropped as it was not properly tested and therefore maintained. I do believe there is value, but it is not a major bottleneck at the moment - 10g from a performance perspective provides a good enough alternative, and latency is still something we should work in the code.
The network stack for example, may be contributing to increased latency.
Removal of translators that are not needed in the stack might help performance as well. We have many translators, but few are needed in most cases.
Y.
On Thu, Jun 11, 2020 at 10:52 AM Federico Strati <strati.federico@xxxxxxxxx> wrote:
Dear All,
I just started working for a company named A3Cube, who produces HPC
supercomputers.
I was assigned the task to investigate which improvements to gluster are
viable
in order to lead to overall better performance in latency and throughput.
I'm quite new to SDS and so pardon me if some questions are naive.
From what I've understood so far, possible bottlenecks are
in FUSE and transport.
Generally speaking, if you have time to just drop me some pointers,
1] FUSE + splice has never been considered (issue closed without real
discussions)
(probably because it conflicts with the general architecture and in
particular
with the write-behind translator)
Recently, it has been announced a new userspace fs kernel module, ZUFS,
whose aim
is to zero copy and improving vastly over FUSE: would you be interested
in investigating it ?
(ZUFS: https://github.com/NetApp/zufs-zuf ;
https://lwn.net/Articles/756625/)
2] Transport over RDMA (Infiniband) has been recently dropped:
may I ask you what considerations have been made ?
3] I would love to hear what you consider real bottlenecks in gluster
right now regarding latency and thruput.
Thanks in advance
Kind regards
Federico
_______________________________________________
Community Meeting Calendar:
Schedule -
Every 2nd and 4th Tuesday at 14:30 IST / 09:00 UTC
Bridge: https://bluejeans.com/441850968
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
https://lists.gluster.org/mailman/listinfo/gluster-devel
_______________________________________________ Community Meeting Calendar: Schedule - Every 2nd and 4th Tuesday at 14:30 IST / 09:00 UTC Bridge: https://bluejeans.com/441850968 Gluster-devel mailing list Gluster-devel@xxxxxxxxxxx https://lists.gluster.org/mailman/listinfo/gluster-devel
