High Load Copying to NFS

I have a large RAID5 array (3x80G) shared via NFS to allow access to photos and music from my other computers on my server. When copying a large number of files over to that system, I’m noticing a huge jump in load. I’m getting a load of 7.5 on the server right now, and I’m just copying files from the client to the server’s NFS share.

The server has the directory shared with this line in /etc/exports


The RAID5 array is defined as

raiddev /dev/md1
raid-level 5
nr-raid-disks 3
nr-spare-disks 0
persistent-superblock 1
chunk-size 32
device /dev/hdf1
raid-disk 0
device /dev/hdg1
raid-disk 1
device /dev/hdh1
raid-disk 2

And the filesystem is mounted in fstab as

/dev/md1 /mnt/share ext3 noatime 0 0

DMA is emabled on the drives, and hdparm -Tt /dev/md1 gives me

Timing buffer-cache reads: 128 MB in 1.38 seconds = 92.75 MB/sec
Timing buffered disk reads: 64 MB in 1.70 seconds = 37.65 MB/sec

I’m also seeing 12mb/s transfer rate through the 100mb hub I have (which isn’t great, but isn’t horrible either πŸ™‚

I’m seeing high load from raid5d, kjournald and nfsd in top and I’m guessing (hoping) that there’s a simple fix for this. It’s not a huge deal, but it’s still annoying. Anyone know why it would be doing this and how to fix it?

7 Comments on “High Load Copying to NFS”

  1. Funnily enough via SMB it’s now a load of 1 or so, and iftop reports 27mb/s speeds.
    Odd, I always thought that nfs was faster. Will have to test more.

  2. I had the same thoughts. But I just did a google search, and found that samba can be tuned better and get more performance.
    Interesting about your load issues being reduced with Samba in that way. Outside of the weakness of SMB over TCP (which you could encapsulate over IPSec if you wanted) seems it might work better for ya. And is native to your windows boxens.
    Go figure. I’m gonna go and tweak my samba server now and kill off the NFS mounts. Well, maybe not. NFS still works better if you want to simply move a dirs/files around on the remote server. Unless it changed recently, Samba still requires a download to the local system and then back up, which is just stupid. Of course… this was a while ago since I had to test that to see if that ever got fixed. Who knows how far the samba team has gotten recently.
    Why don’t you test and let us know πŸ™‚ Copy a huge directory from one sub to another on the remote system.. and see which is faster. πŸ™‚

  3. I’m going to go out a limb here and guess that you aren’t using a Sun server on your network. If that’s true I ain’t got nothing for you πŸ™‚

  4. Dana – it’s not instant, but it’s far faster than moving to the local box and back again. Easy enough to try yourself though πŸ™‚
    IPSEC is nice, but really not an issue in my LAN where it’s only me and A. I doubt she’ll be sniffing the data, as she can mount the shares herself πŸ™‚ Nifty docs though, will have to do some more reading….
    Raskal – no, sadly no suns here.

  5. dude, an 80 gig RAID5 is not large by any stretch. When you get into multiple terabytes, call me πŸ˜‰

  6. Engel – it’s actually 160G, but still, you’re right, it’s tiny. Course, it’s also only got 28 G left on it, and when I’m done todays trip to alt.binaries.mp3.* it’ll have far less. Time to get 3 or 4 120s or 250s….