I am only getting write transfer speeds of about 16-17MB/s over my gigabit network.  This seems low but maybe I had too high expectations.  I was expecting something in the range of 20-30MB/s.  The speed is the same whether writing to the server with either XP or Leopard 10.5.1 (both laptops have gigabit NICs)

 

 

Server Specs

Motherboard:  Asus P5B-VM DO

CPU:  P4 3.0GHZ

Memory:  2x 512MB

Hard drives:  3x SATA 750GB Samsung Spinpoint.  Note: I have enabled "Joe L's 30% performance tweak" and my hard drive test yields 86.3MB/s

Network switch:  Netgear GS605

 

 

So my questions to the forum are:  What network throughput do you guys get while writing files (1GB+ in size) to your shares?  Should I be getting faster write throughput than 16-17MB/s given that I used relatively higher end server components?

 

 

 

that speed is normal for writes, i get about 15-20/MB writes, and Tom has stated that you won't get much more than that.

Since yours is a newly built server you could not assign a parity drive at the moment, transfer over all your data, and then assign a parity and let it build the parity...it will speed up your trasnfers...but as always, be careful

I try to push all my "big" updates overnight... just queue em up and dump em all at the same time and goto bed.

That's about the speed I get.  My understanding is every write operation has to become a read-modify-write on the parity drive to calculate the new parity, plus a write on the data drive of course.

 

I think the only way to avoid that would be to do read operations on all other drives, and a write on the parity & data drive, and that has the disadvantage of forcing all the drives to spin up.  Probably also a pain in the butt to implement efficiently too.

 

 

 

 

My understanding is every write operation has to become a read-modify-write on the parity drive to calculate the new parity, plus a write on the data drive of course.

 

Worse!  It's a read and write on both the parity drive AND the data drive.  It has to know exactly what bits are changed in the data block before it can change the corresponding parity block.  If I understand right, Tom has said he is making some optimizations of this process, that should increase the write speed.

 

Ditto.  Bananafish I've got the same motherboard and samsung spinpoints as well (500's).  When writing I generally get about 1G per minute, which works out to about 16.7 Mbps transfer.

Everyone is really talking about MBps, right???  16-17 Mbps on GigE, even for writes, would be truly awful...

I hate the whole MBps vs Mbps thing.  It would be a lot simpler if people just spelled out what they were trying to say Megabit or Megabyte. 

 

Its important for people to realize that you never actually get the speed that any network connection is rated at.  If you are getting 16 or 17 Megabytes per second sustained, on home networking equipment.  Using computer equipment meant for home use, thats actually pretty darn good. 

 

Occasionally I'll see spikes of 200 Megabits per second on my setup, but I think to really get much faster than that you'd have to use real server class hardware, connected with real enterprise class switching gear. 

 

If you've got a pretty peppy processor, I don't see why parity calculations would slow network transfers down that much. 

It is not the parity calcs that slow down writing to the unRaid array, but that you must read both the block you are writing to , and the parity block you are going to replace before you then write them.

 

So, instead of a single write operation, you have two read operations, and two write operations, just to write a single block to any disk.  If not properly buffered, you mist wait for the disks to spin around multiple times before you can write the same block on the disk you just finished reading. 

 

For RAID arrays to be reasonably reliable, you cannot do too much buffering in memory of blocks written, otherwise the array is not valid if you lose power before the memory contents are written to the physical disk.  It is a trade-off of performance vs. physical writes and reliability.

 

Now, other issues are at work too.  We have seen in earlier releases that network IO was stopped when the background process is writing to the physical disks and the buffering of writes was full.  That will certainly slow network traffic, or at least make it appear to occur in bursts.

 

Joe L.

TSM, what makes you think that spelling it out would make a difference?  If someone doesn't know what it means, they are just as likely to say "megabit" instead of "megabyte"--in fact I've seen that as often as the Mbps vs MBps mixup... 

 

And I simply can't agree that 16-17MBps is "pretty darn good" on Gigabit.  Although unRAID might be limited to ~17MBps writes (with some good reasons, it seems), even consumer-grade gigE equipment has several times that in headroom.  Just assuming a maximum of 50% utilization (frankly, even 70% is cautious) would yield over 60MBps.  No matter how you slice it, at 17MBps the slow link is not the network--even for low-end equipment...

Everyone is really talking about MBps, right???  16-17 Mbps on GigE, even for writes, would be truly awful...

 

Yeah sorry.  I'm talking about MB.  I think everyone understood but I modified my original post anyway.  Thanks

 

Glad to hear there's nothing wrong with my 16-17MB/s performance.  Overall i'm extremely happy I went with unRAID.

 

 

...

Its important for people to realize that you never actually get the speed that any network connection is rated at.  If you are getting 16 or 17 Megabytes per second sustained, on home networking equipment.  Using computer equipment meant for home use, thats actually pretty darn good. 

...

 

Hmm - I'd have to disagree.  I only buy cheap network gear (linksys/netgear switches, etc) and I've done way, way better than that.

 

My experience is with wired ethernet, you can get pretty close to the rated speed unless something else is slowing it down.  I'm pretty sure I've exceeded 40MB/sec going to either my old raid-1 box or through windows file sharing, though I can't remember which one. And I definitely get better than that with reads to the unRaid box.

 

I haven't got hard drives that can go as fast as gigabit ethernet, though I've seen some pretty good rates using bing to test network throughput - from my laptop I get ~75MB/sec to my unRaid box, ~100MB/sec to my desktop box.  That's getting past 80% efficiency.

 

Now wireless is a different story - the numbers thrown above are WELL beyond what you actually get.  Even with 802.11n on the 5g band, I can only sustain 8-9MB/sec - nowhere near the 270Mb/sec (33MB/sec) that's advertised.

...

Its important for people to realize that you never actually get the speed that any network connection is rated at.  If you are getting 16 or 17 Megabytes per second sustained, on home networking equipment.  Using computer equipment meant for home use, thats actually pretty darn good. 

...

 

Hmm - I'd have to disagree.  I only buy cheap network gear (linksys/netgear switches, etc) and I've done way, way better than that.

 

My experience is with wired ethernet, you can get pretty close to the rated speed unless something else is slowing it down.  I'm pretty sure I've exceeded 40MB/sec going to either my old raid-1 box or through windows file sharing, though I can't remember which one. And I definitely get better than that with reads to the unRaid box.

 

I haven't got hard drives that can go as fast as gigabit ethernet, though I've seen some pretty good rates using bing to test network throughput - from my laptop I get ~75MB/sec to my unRaid box, ~100MB/sec to my desktop box.  That's getting past 80% efficiency.

 

Now wireless is a different story - the numbers thrown above are WELL beyond what you actually get.  Even with 802.11n on the 5g band, I can only sustain 8-9MB/sec - nowhere near the 270Mb/sec (33MB/sec) that's advertised.

 

OK, just so I'm clear.  You get 800 megabits throughput?

 

OK, just so I'm clear.  You get 800 megabits throughput?

 

On a somewhat synthetic test, yes.  This is basically sending large ICMP packets, not anything that involves data coming from or going to a disk.

 

I've seen rates of ~500megabit sustained doing some samba reads, but that seems to be when the data is cached on the unRaid box.  When reading a newly-accessed file, the rates are more like 250megabit sustained.

 

If you want, I can make a static-linked binary of bing to play with.  If you get rates  <50% utilization, I'd check and see if you're getting any errors or collisisons on your ethernet interfaces.

I get around 320Mbps sustained (40MBps) when copying from unRAID to my somehow slow MacBook and Mac Mini hard-drive.

And in the 10-20 MBps fluctuating range when writing to unRAID.

 

 

 

how are you guys testing this?  is it a guess?  or tool?