GlusterFS 8+: Gluster Roadmap

I/O Performance:

Performance of ls -l command

Especially when things are not ideal, self-heal is pending, when there are activities etc.. etc..

• What are observations?
• How to fix it?
• Any design changes?
  https://github.com/gluster/glusterfs/issues/611
• RCA and fix the regression in parallel-readdir as observed by Nithya

Network Layer

• Challenges with current n/w layer.
  https://github.com/gluster/glusterfs/issues/391
• Rpc being bottleneck with nvme backend
  https://bugzilla.redhat.com/show_bug.cgi?id=1467614#c52
  • We don’t have an RCA on why single rpc connection is underperforming. One observation from Krutika is that reading a single rpc msg can lead to utmost 4 reads on socket. I’ll experiment whether reading a single msg with lesser number of reads increase performance
  • Note that the bug is closed because we made some improvements in the area, soon will open a bug with more details.

Performance Group Configuration files

• There are no volume option groups defined to prepare a volume for a special workload e.g. low latency, small files, big files, etc
• According to Redhat it is easy to create such groups, no change to the code has been made:
  https://access.redhat.com/documentation/en-us/red_hat_gluster_storage/3.3/html/administration_guide/configuring_multiple_volume_options
• Having groups for specific workloads would be a benefit for the user because at the moment there are so much volume options and it is hard to understand
  them all and the way they are affects each other. So a well documented group would be easy to understand and easy to use

Monitoring / Instrumentation

• How do we understand what is causing latency in a given fop?
  If a user / customer reports a performance issue, how do we approach the problem?
  https://github.com/gluster/glusterfs/issues/688
• Some thoughts on adding metrics to caching layer can be found here

Caching Layers

• Fuse Write-back
• Validate current stack
  • Some pointers on this at https://www.usenix.org/conference/vault19/presentation/pillai
  • Read-ahead and io-cache can be removed from default xlator stack on fuse mounts - https://review.gluster.org/22203

Other issues

• https://github.com/gluster/glusterfs/labels/FA%3A Performance Improvements

Management Layer Performance:

Improvements on static scan of volume list

With container native storage, when we deal with thousands of volumes having scanning a list based volume entries to find out a volume object isn’t efficient. We’re looking to optimize this to have a better performant method.

• Initial idea : https://review.gluster.org/#/c/22938/

Scale Improvements

RIO feature

• More on it @ https://github.com/gluster/glusterfs/issues/243

• With the current 'known' limitation of directory scale of DHT model, the best solution ahead for gluster project is to get RIO to completion and have a solutions for large scale deployments.

Other improvements

• https://github.com/gluster/glusterfs/labels/FA%3A Scalability Improvements

Container Story

We see improved usage of Gluster in container solutions for getting persistent volumes (PV & PVC). Let us know what are the options you have considered, and what is good for you etc etc.

heketi

• https://github.com/heketi/heketi

gluster-subvol

• https://github.com/gluster/gluster-subvol

something-else?

• https://devconfin19.sched.com/event/RVPw

Other topics

Path based geo-replication

• https://lists.gluster.org/pipermail/gluster-devel/2019-February/055836.html
  • https://docs.google.com/document/d/1gW5ETQxNiy9tt4uV1ohRH1g5AMmWLtbQYD3QPs_v8Ec/edit
• https://github.com/gluster/glusterfs/issues/460

Improvement to testing ?

• How to get results faster?
• Are we using our resources better?