CPU usage

[brucifer]

Hello,
I am looking for suggestions to improve performance with my current iSCSI setup.
My main use for starwind is to serve disk images to a few video editing workstations. I am not using any HA or virtual machines.
I am using an adaptec 52445 with 28 direct attached 3TB drives in a RAID 50 configuration. This yields very good performance
using BlackMagic Disk Speed Test reading and writing ~800MB/sec. I am running Windows Server 2008 x64 with 48GB of Ram.
I have dual E5540's at 2.53GHz. I have a myricom 10Gb CX4 to a HP 10Gb CX4 switch. My Myricom has: Flow Control On, Large Send Offload Enabled,
MTU 9000, Receive Buffers of 2048, TCP Checksum Offload Enabled.

I am noticing that when performing file copy operations to the starwind targets I see great performance, but only for a short while.

I notice that on the starwind server machine I am seeing that 1 cpu core is pinned at 100% while the others are barely
even being used. If I wait until the copy is finished the single cpu core that is at 100% drops off to almost 0% usage again. When I try another copy
the same thing happens, copy speeds are great, then the cpu usage on a single core ramps up to 100% and then copy speeds drop down significantly.
I notice that the single cpu core that is pinning does change to other cores and act's the same.
Is there anything I can do to spread out this cpu usage or any other suggestions to make this work better?
thanks
brucifer

[anton (staff)]

Current architecture of StarWind involves many threads and we have to bond them to the same CPU to avoid NUMA issues so we're basically limited with the performance of the single CPU. We're working on redesign but it will take time so treat it as "known issue" (thanks for reporting either way). However server should not experience any "hiccups" and you may want to watch that and maybe ask our engineers to jump in for remote session @ some point. Back to blazing fast performance and then slow down: if you don't use LSFS that plays nice with parity and stripped RAIDs - it's normal (you use RAID50 which is both parity and stripe). See you fill the write back cache (that's where performance is high and steady) and then lazy writer starts flushing data to disk and tons of random writes kill parity RAID because "read-modify-write" sequence. So either use RAID10 that's immune to it or use LSFS.