Hi,
Apologies for the late reply. My email filters are messed up, I missed reading this.
Answers to questions around shard algorithm inline ...
On Sun, Sep 30, 2018 at 9:54 PM Ashayam Gupta <ashayam.gupta@xxxxxxxxxxxxxx> wrote:
Hi Pranith,Thanks for you reply, it would be helpful if you can please help us with the following issues with respect to sharding.The gluster version we are using is glusterfs 4.1.4 on Ubuntu 18.04.1 LTS
- Shards-Creation Algo: We were interested in understanding the way in which shards are distributed across bricks and nodes, is it Round-Robin or some other algo and can we change this mechanism using some config file.
E.g. If we have 2 nodes with each nodes having 2 bricks , with a total of 4 (2*2) bricks how will the shards be distributed, will it be always even distribution?(Volume type in this case is plain)- Sharding+Distributed-Volume: Currently we are using plain volume with sharding enabled and we do not see even distribution of shards across bricks .Can we use sharding with distributed volume to achieve evenly and better distribution of shards? Would be helpful if you can suggest the most efficient way of using sharding , our goal is to have a evenly distributed file system(we have large files hence using sharding) and we are not concerned with replication as of now.
I think Raghavendra already answered the two questions above.
- Shard-Block-Size: In case we change the features.shard-block-size value from X -> Y after lots of data has been populated , how does this affect the existing shards are they auto corrected as per the new size or do we need to run some commdands to get this done or is this even recommended to do the change?
Existing files will retain their shard-block-size. shard-block-size is a property of a file that is set at the time of creation of the file (in the form of an extended attribute "trusted.glusterfs.shard.block-size") and remains same through the lifetime of the file.
If you want the shard-block-size to be changed across these files, you'll need to perform either of the two steps below:
1. move the existing files to a local fs from your glusterfs volume and then move them back into the volume.
2. copy the existing files into a temporary filenames on the same volume and rename them back to their original names.
In our tests wrt vm store workload, we've found 64MB shard-block-size to be good fit for both IO and self-heal performance.
- Rebalance-Shard: As per the docs whenever we add new server/node to the existing gluster we need to run Rebalance command, we would like to know if there are any known issues for re-balancing with sharding enabled.
We did find some shard-dht inter-op issues in rebalance in the past again in the supported vm storage use-case. The good news is that the problems known to us have been fixed, but their validation is still pending.
We would highly appreciate if you can point us to the latest sharding docs, we tried to search but could not find better than this https://staged-gluster-docs.readthedocs.io/en/release3.7.0beta1/Features/shard/ .
The doc is still valid (except for minor changes in the To-Do list at the bottom). But I agree, the answers to all of the questions you asked above are well worth documenting. I'll fix this. Thanks for the feedback.
Let us know if you have any more questions or if you run into any problems. Happy to help.
Also, since you're using a non-vm storage use case, I'd suggest that you try shard on a test cluster first before even putting it into production. :)
-Krutika
Thanks
AshayamOn Thu, Sep 20, 2018 at 7:47 PM Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx> wrote:On Wed, Sep 19, 2018 at 11:37 AM Ashayam Gupta <ashayam.gupta@xxxxxxxxxxxxxx> wrote:Please find our workload details as requested by you :* Only 1 write-mount point as of now* Read-Mount : Since we auto-scale our machines this can be as big as 300-400 machines during peak times* >" multiple concurrent reads means that Reads will not happen until the file is completely written to" Yes , in our current scenario we can ensure that indeed this is the case.But when you say it only supports single writer workload we would like to understand the following scenarios with respect to multiple writers and the current behaviour of glusterfs with sharding
- Multiple Writer writes to different files
When I say multiple writers, I mean multiple mounts. Since you were saying earlier there is only one mount which does all writes, everything should work as expected.
- Multiple Writer writes to same file
- they write to same file but different shards of same file
- they write to same file (no gurantee if they write to different shards)
As long as the above happens from same mount, things should be fine. Otherwise there could be problems.There might be some more cases which are known to you , would be helpful if you can describe us about those scenarios as well or may point us to the relevant documents.Also it would be helpful if you can suggest the most stable version of glusterfs with sharding feature to use , since we would like to use this in production.It has been stable for a while, so use any of the latest maintained releases like 3.12.x or 4.1.xAs I was mentioning already, sharding is mainly tested with VM/gluster-block workloads. So there could be some corner cases with single writer workload which we never ran into for the VM/block workloads we test. But you may run into them. Do let us know and we can take a look if you find something out of the ordinary. What I would suggest is to use one of the maintained releases and run the workloads you have for some time to test things out, once you feel confident, you can put it in production.HTHThanksAshayam GuptaOn Tue, Sep 18, 2018 at 11:00 AM Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx> wrote:On Mon, Sep 17, 2018 at 4:14 AM Ashayam Gupta <ashayam.gupta@xxxxxxxxxxxxxx> wrote:Hi All,We are currently using glusterfs for storing large files with write-once and multiple concurrent reads, and were interested in understanding one of the features of glusterfs called sharding for our use case.
So far from the talk given by the developer [https://www.youtube.com/watch?v=aAlLy9k65Gw] and the git issue [https://github.com/gluster/glusterfs/issues/290] , we know that it was developed for large VM images as use case and the second link does talk about a more general purpose usage , but we are not clear if there are some issues if used for non-VM image large files [which is the use case for us].Therefore it would be helpful if we can have some pointers or more information about the more general use-case scenario for sharding and any shortcomings if any , in case we use it for our scenario which is non-VM large files with write-once and multiple concurrent reads.Also it would be very helpful if you can suggest the best approach/settings for our use case scenario.Sharding is developed for Big file usecases and at the moment only supports single writer workload. I also added the maintainers for sharding to the thread. May be giving a bit of detail about access pattern w.r.t. number of mounts that are used for writing/reading would be helpful. I am assuming write-once and multiple concurrent reads means that Reads will not happen until the file is completely written to. Could you explain a bit more about the workload?_______________________________________________ThanksAshayam Gupta
Gluster-users mailing list
Gluster-users@xxxxxxxxxxx
https://lists.gluster.org/mailman/listinfo/gluster-users
--Pranith
--Pranith
_______________________________________________ Gluster-users mailing list Gluster-users@xxxxxxxxxxx https://lists.gluster.org/mailman/listinfo/gluster-users
