2 thoughts on “Giddy …

  1. Looking forward to it. Bull had some interesting boxes that got *very* good performance at LUG using dual IO hub configurations. Unfortunately it required patching Lustre to issue OSS service threads to cores on the socket associated with the IOH servicing the raid controller for the OST being targeted. I don’t know enough about how Gluster does reads/writes behind the scenes, but if the NUMA topology can be taken into account some pretty insane performance can be realized.

  2. @ Mark

    This isn’t a Gluster box. Theoretical max performance for write, for the single machine is north (far north) of 10GB/s.

    Measured is a different story. Working on understanding and discovering fun limits/interactions of PCI, drivers, etc. along the way.

    Cursory IOP measurement, doing nothing special was 120k for 8k random writes. Should have been much higher … theoretical max much closer to PCIe limits.

    Cursory BW measurement for sequential reads/writes had some nice peaks for a while at 5+ GB/s, then we averaged in the mid/upper 3 GB/s range.

    Part of this is RAID layout, part is RAID chip performance. Lots of it are limitations due to the HBA design and driver stack implementation. Some tweaking today got the performance up to 6 GB/s sustained for reads. Writes are still too slow IMO (3.7 GB/s, should be much better than that).

Comments are closed.