Forums

Home / Forums

You need to log in to create posts and topics. Login · Register

CSVFS volumes slow while fast when not assigned as cluster volume

Hello everybody,

any ideas on why iSCSI LUN is quite fast when assigned as "standard" iSCSI directly to a Win Srv 2019 node while getting incredebily slow when assigned as CSVFS?

Setup:

  • 5 nodes (3 mgm, 2 stor) each with 3x10T HDD and 1x1T SSD as cache
  • 2x40G network between nodes
  • 4x10G network between nodes and Windows server(s)

Benchmarks:

  • IO write performance when iSCSI lun is "directly" attached as NTFS to a node: ~900 MByte/s
  • IO write performance when iSCSI lun is used as CSVFS: ~50 MByte/s

Can't see any bottlenecks at CPU nor RAM nor network.

Any help or suggestions are strongly appreciated.

Have looked at our Hyper-v and/or SFOS and MPIO guides in setting up CSV ?

Are you using CSV on top of NTFS ?

How you test this ? Do you have CrystalDiskMark ? if so can you show irs results ?

Hi,
1.) CrystalDiskMark (when running to "just" attached as iSCSI lun formatted with NTFS, not attached to cluster):

  • Sequental 4M Q8T10 Mix 1105.47
  • Sequental 4M Q1T10 Mix 1059,48
  • Random 64K Q32T10 Mix 1122.59
  • Random 64K Q1T10 Mix 419,90

2) Yes, we do - the iSCSI luns are attached to Hyper-V Failover Cluster; as we do for our "primary" PetaSAN cluster

3) Did several tests:

  • Running CrystalDiskMark as described above
  • Copying file(s) - bunch of small ones as well as few huge ones (>5G) - from node local storage to the new one as "one thread", copying files "multithreaded" (robocopy MT-option)
  • Copying file(s) as above from "primary" PetaSAN cluster to the new one
  • It makes no difference whether the tests are run on the node, which is currently in charge of the LUN or on another node - the performance difference is negligible (55 <-> 50 MB/s)

Can you show the CrystalDiskmark for the CSV case

Also if you can show 4k Q1T1 test for both cases

Nevermind - sorry for the (unnecessary) efforts ... it wasn't PetaSANs fault but the way CSVFS works (= long story short).