I/O Scheduler

[lightp2]

Side note about Linux I/O experience

1. One of the work PCs runs 4-disk Linux MD RAID-10.
2. I ran some simple benches and everything reasonable for single user.
3. This host exports NFS to VMWARE for VM Storage.

On this particular occasion,
4. only one test Windows VM doing dir c:\ /s
5. The Linux desktop response became sluggish.
6. I was surprised because it should have no problem with such light load.
7. iostat - IOPS wasn't extreme.

Finally I did the necessary adjustment for all the drives, example for one of the drives

echo deadline > /sys/block/sda/queue/scheduler

8. Now everything back to reasonable.

My Tentative experience.

I think the default I/O Scheduler (CFQ) maybe I need more understanding and tuning for non-desktop cases. For me at least, deadline I/O Scheduler seems to be reasonable in most cases. I am not a top performance seeker.

 

[Eva_Unit_0]

I've always liked the deadline scheduler and I've personally felt that CFQ has historically given mixed results (though it has seen some major improvements recently). Anyway, when you're dealing with storage that is abstracted (like your MDRAID over NFS), using the plain no-op (noop) scheduler is often the best choice. This is because 1. it becomes very difficult for the scheduler to predict usage patterns and 2. often it is impossible for the scheduler to actually do any optimization anyway. Try using "noop" and see how it works.

Alternatively, is there a hardware problem of some type? Are you seeing a large amount of hardware interrupt load (HI% in top)?

 

[lightp2]

Hi, my apology for late reply as I missed your post earlier.

1. During the single VM dir c:\ /s test, on the NFS server, top %hi between 0.9 to 2.2
I have no exact idea about whether this is OK or not. I RDP into the test VM from the Linux NFS Server to perform the test and found the screen GUI partially hung with no response while the console texts flew by.

2. I also tested noop scheduler as recommended and found it performed as described with no noticeable lag. However, I found the top %hi is slightly higher between 3-4.7%

3. Currently I still use deadline I/O scheduler as I am also using PC as normal desktop in addition to NFS server. I have run some seeker tests with various system settings modification and found deadline slightly better most of the time.