Gigabit tuning & performance - what do you get?

[deed02392]

Hi guys, I know this has been discussed to death but bear with me!

My FreeNAS box is using 3 2 TB disks in a RAID-Z1 arrangement, with the default CIFS settings set up and an export of the whole ZFS volume set up.

I am also running the ProFTPd, with no encryption on data channels.

From a wired gigabit ethernet Windows 7 machine, I can't seem to get better than about 70 MByte/s at best (both via FTP transfers and Samba). This is much slower than transfers I get between directories directly on the system itself (`cp`). But I suppose that is to be expected. Since gigabit's limit is around 110 MByte/s, I expect I should be getting at least 90% of that accounting for typical overheads, that being more like 95 MByte/s. Yet typical transfers are 60 MByte/s.

Anyway, what measure should I look into for increasing performance on all clients? The network card on the server is a high-end HP NIC, supporting jumbo frames and iSCSI etc, although I don't think I utilise any of these at the moment. The NIC in my wired desktop machine is just a generic onboard one that came with my ASUS mobo. My switch is a Linksys E3000 running Tomato-usb.

Server specs:

Code:

16 GiB DDR3 1033 MHz
Celeron 847 Dual-core 1.1 GHz
3x 2 TB 7200 RPM (RAID-Z1)
HP NC380T 374443-001 Dual-port GbE

I'd like to get a discussion going regarding what speeds people here are getting and what their configurations are. Perhaps we can come up with some optimal configurations for a typical GbE wired network.

Edit: Interestingly, I just copied a file from a different hard drive than I normaly would and the transfer was 100 Mbyte/s... why does one always do things like this after they've written up a forum thread. So perhaps the local HDD is quite fragmented or simply slow!

 

[jgreco]

You should run iperf and see what sort of speeds you get. By cutting out all the filesystem cr*p and protocol cr*p, you can focus on testing just one little portion of the problem that MIGHT be a problem or it might NOT. When you start talking using your access-point/NAT-gateway as a switch that you expect peak performance out of, and you are using a generic desktop ethernet interface as one of your endpoints, there are multiple opportunities for fail that are outside the realm of FreeNAS itself. That's why a quick test is in order.

 

[deed02392]

Just ran iperf and get 940 Mbit/s between the two boxes. But I am getting that level of performance out of even Samba when I use a fast enough disk, so I guess that must have been my issue - I just assumed all HDDs had a throughput of at least 100 MByte/s but I don't think that's really always the case.

 

[jgreco]

Oh no, many new hard drives are more than 100MB/sec, but plenty of older drives are not. The fastest drives around 2005 capped out well below 100MB/sec. Also, any speed you get like that is only for _sequential_ access; random access is ALWAYS slower due to the need to reposition heads, which can only happen very slowly. So if you're reading lots of small files or seeking all over the place, slowness.

 

[deed02392]

Yeah I can appreciate that. I think everything is good here in that case. Just wish I had 10 GbE so I could really test out this bad boy.