ZFS high io. Again...
- Thread starter Kot Degt
- Start date
Noop is like no schedule op. Deadline it is a schedule io because have 2 different queue, one for write op and one for read op. 60 % of the total queue size is for reading io and 40% for write io. Also if I remember each io op has a max time to wait until is sent to the disk (so the name is deadline)
Now it is true that you can use in zfs a entire disk, and by default hdd schedule is noop. But in reality zfs can use only partitions for a good reson.
Now it is true that you can use in zfs a entire disk, and by default hdd schedule is noop. But in reality zfs can use only partitions for a good reson.
I'm experiencing huge CPU loads when doing sequencial read/write. The server becomes unresponsive until the transfer ends.
My setup:
2x Xeon E5 2620 v3
96GB RAM
12x 2TB SAS 7200RPM in RAIDZ2
1x ZIL
1X L2ARC
4x 1Gbps in LACP
I can do various bonnie++ benchmarks with outstanding results (~600 MB/s writes and about 1.6 GB/s reads)
Inside my VMs I get very good random I/O benchmarks aswel.
But...
When I try to copy a large file from another server into a VM i get a few seconds of 1Gbps transfers than it plumets and the server load sky rockets.
(Don't mind the gaps, I was testing various ZFS parameters)
My setup:
2x Xeon E5 2620 v3
96GB RAM
12x 2TB SAS 7200RPM in RAIDZ2
1x ZIL
1X L2ARC
4x 1Gbps in LACP
I can do various bonnie++ benchmarks with outstanding results (~600 MB/s writes and about 1.6 GB/s reads)
Inside my VMs I get very good random I/O benchmarks aswel.
But...
When I try to copy a large file from another server into a VM i get a few seconds of 1Gbps transfers than it plumets and the server load sky rockets.
(Don't mind the gaps, I was testing various ZFS parameters)
It is not. I can transferir 2 streams of sequential data at 100MB/s without a sweat in that "another server".
Doing transfers between VMs generate the same absurd high loads.