0

I'm having an infrastructure design problem with an application. Right now we access a lot of small files (under 10MB) from a GFSv2 2-node cluster. 90% of file access is "random reads" to this GFSv2 partition the other 10% is random writes. I've done all the tuning for noatime, nodirtime, and plocks but the IOwait is still too high. What are the better alternatives for this type of scenario?

Other possibly relevant details: all gigabit networking, all hosts are in the same rack, gfs comes from a SSD tiered SAN with <1ms latency and great performance, iowait with DLM in-use is 3% just writing two 3MB files per second. We obviously plan to handle a lot more than this. I need a solution that will be HA and scale across horizontally.

I know filesystem choice is very dependent upon the type of traffic so I hope I've given an accurate depiction of my use-case.

onetruescottsman
  • 1
  • 1

1 Answers1

1

It depends on how small the files are. If you are over 10kb you can try GlusterFS. It is great with bigger files and 2 node cluster with mirror should be able to handle a lot of throughput. With really small files - gluster fails :(

You can also try Ceph (object or block storage) or Swift (object). The thing with object storage is that you need to connect to it with api.

piotrektt
  • 184
  • 2
  • 14