Gigabit speed tests performed. Who can help analyze the results

[Jeroen1000]

Hi guys,

I'll try to be brief and to the point here so that I don't make a mess of the opening post.

Test System 1:

XP System:
Processor: Phenom II X3 720
Ram: OCZ 4GB DDR3 1333Mhz Reaper HPC Matched Pair
HD: Samsung 1Tb SATA II 32Mb HD103UJ
Network card: Atheros AR8121, AR8113, AR8114, PCI-E ethernet controller
OS: Windows XP 32-bit service pack 3
Firewall and AV off. Forgot to disable Windows Defender.

Test System 2:

Vista System:
Processor: Intel Core 2 Duo P7350
Ram: Hynix (4GB)
HD: WDC WD3200BEVT-22ZCT0
Netword card: Intel ( R) 82567LF Gigabit Network Connection
OS: Windows Vista home premium 32-bit service pack 2
Firewall and AV off. Forgot to disable Windows Defender.

Network equipment:

PROSAFE® 5-PORT GIGABIT ETHERNET DESKTOP SWITCH 10/100/1000 MBPS (link: http://www.netgear.com/Pr...spx?detail=Specifications)

Software used to test:

- Mozilla FTP- client and server
- DataRam ramdisk (http://memory.dataram.com...services/software/ramdisk)

Remarks:
- I have tested both CIFS and FTP protocols
- I have used cat5e cables from a computer store. I have not patched the cables myself. Both cables used were 1.5 metres in lenght.
- I have completed tests with and without a ramdrive. If you check out the Excel sheet, "Met RAMDRIVE" is Dutch for "with RAMDRIVE". "Zonder RAMDRIVE" is Dutch for "without RAMDRIVE". If a test was performed with a RAMDRIVE, both computers were using one.
- The file that was transfered was exactly 1 gigabyte, no more no less.
- I am aware a rounding error in the sheet exists. I'll correct it later. The corrected version is at home. The error is insignificantly small.

Link to Excel sheet: http://www.megafileupload.com/en/file/194444/Gigabit-testen-xls.html

Regarding the ramdrive:
There are a few things I would like to ask that more network savvy people may be able to answer. My ramdrive has a bandwidth exeeding 3 gigabyte per second. So we can all safely agree it is not the bottleneck. However, I have performed a test on it with rather odd results.

I have written to the ramdrive using a tool called HD speed. When I continuously write a 1 kilobyte file (1024 bytes), I achieve a recorded speed of only 51 megabytes per second.
A 2 kilobyte file yields a speed of 101 megabytes (= 105906176 bytes) per second. A 4 kilobyte file gets me close to 200 megabytes per second (about 195 to be more accurate).

Still with me here?

So IF an ethernet frame is 1518 bytes and a 2048 byte file yields a speed of 105906176 bytes per second, my theoretical max. speed would be a mere 75 megabytes per second. Of course, when you check out the Excel, you can see FTP-got me close to 90 megabytes per second on average.

Can anyone explain these results?

Now, obviously there is still a bottleneck. Is it the OS? Is it the ramdrive (probably not the drive) or perhaps it is the ethernet controller or its driver? I'm not excluding FTP- protocol overhead but only 90 megabytes/second from ramdrive to ramdrive is more than just 'protocol overhead'.

Any advice would be very welcome.

Cheers,
Jeroen

 

[thiggins]

Internal drive speeds are one thing, network transfer speeds quite another.

Assuming that your RAM drive is not a limitation, block sizes used in the network transfers will limit your speed. For CIFS transfers, XP is probably limiting your speed. If you used two Vista SP1 or higher systems, you'd see your speed increase, again if the RAM drives aren't the bottleneck.

See
How To Build a Really Fast NAS - Part 6: The Vista (SP1) Difference

 

[Jeroen1000]

I've installed Windows 7 yesterday (on the XP computer, which is now dual booting with Win7). I've ran several Iperf tests with 64k windows (the default window was too small). Both the Win7 desktop and the Vista laptop were directly connected (with short cables). I tested 2 cable sets. Iperf gave me about 500 megabits/second in all cases: Iperf on vista and xp, Iperf on vista and Win7, iperf on xp and win 7. All 'round the 500 megabit/s mark. Grrr!

Subsequently, I did a CIFS/Samba transfer from a Vista WITH ramdrive to Win7 without Ramdrive. Speed got limited to 62.2 megabytes per second, which is disturbingly close to what Iperf had reported earlier. I'll try a ramdrive to ramdrive transfer between Vista and Win 7 soon, but I was too lazy yesterday to hook up the internet to download the ramdrive tool again as it was getting late.

I'm convinced the ramdrives are good. However, why am I getting such slow Iperf results. Tim, would you reckon using Xia chariot would produce some more reliable results? Or could you please suggest another test suite?

 

[Jeroen1000]

I just used Iperf between 2 office computers. I connected them via a 2 metres CAT5e UTP cable. Result 900 to 919 megabit per second.

Yeah, they have Intel pro nic's and the systems are almost identical They both run XP SP3. Darn, what's wrong with my systems

 

[thiggins]

Shut off the Windows firewall if you have it running. Poke through services and see if there is something network related that you can kill.

Check the NIC Connection properties and uncheck anything except for
TCP/IP and Client for Microsoft Networks.

I'm assuming you've checked Windows task manager for CPU and Network utilization.

 

[00Roush]

Hi guys,

I'll try to be brief and to the point here so that I don't make a mess of the opening post.

<BIG SNIP!!!>

Maybe I can help explain your results a bit. First you need to realize that the smaller read/write sizes you used with HD speed are not actually what windows generally uses to read and write files with. With XP common sizes used were 4k and 64k. (other sizes are used but I believe these are the most common) 64k was the largest size XP could use. Newer windows versions can use larger sizes but I believe 4k and 64k are still commonly used. The reason you need to know this is because the smaller the read/write size the more overhead you get. For example you want to write a 1024k file to the disk. You can either do 1024 writes at 1k or you could do 16 writes at 64k. The latter has considerably lower overhead and can be processed much quicker.

Now you mentioned ethernet frame sizes. You are correct that they are fairly small but this is not the actual size that is used when reading and writing to the filesystem. In general windows will actually try to read/write from/to a remote file just like it would a local file. It does this via SMB. (CIFS/Samba) Ethernet is just what is used to transport the data to and from each computer. So the actual read/write sizes from/to the disk turn out to be similar to what I mentioned above.

I hope that helps a little.

00Roush

 

[Jeroen1000]

Shut off the Windows firewall if you have it running. Poke through services and see if there is something network related that you can kill.

Check the NIC Connection properties and uncheck anything except for
TCP/IP and Client for Microsoft Networks.

I'm assuming you've checked Windows task manager for CPU and Network utilization.

I did disable the firewall and eventually also Windows defender. That didn't do me any good. I did not check, however, the task manager. CPU-utilization was never an issue but I could have checked the network utilization. In the end I decided to update the ethernet controller driver. There weren't any via the Fujitsu website (for the laptop) as I had already checked. But on Intel's website there were!
Now, I'm getting about 800 megabits per second (via Iperf). Good enough for me!

 

[Jeroen1000]

@OORoush, I assume you are saying that the OS buffers the data until it has enough to write at least 4k at a time? There are some buffer related settings in the drivers so I think that takes care of itself. I kind of made a wrong assumption using HD-speed.

 

[00Roush]

@OORoush, I assume you are saying that the OS buffers the data until it has enough to write at least 4k at a time? There are some buffer related settings in the drivers so I think that takes care of itself. I kind of made a wrong assumption using HD-speed.

Yea sort of... I suppose what I really was trying to relay was that writing or reading lots of small files (smaller than 4k) takes a lot of overhead and therefore decreases performance. This goes for just about any storage device. As was pointed out by your HD-speed program results with your ramdrive.

I would recommend testing with 16k, 32k, or 64k if you are looking for the performance of a particular drive. These larger read/write sizes are used more often by the OS when transferring larger files and therefore would give a better indication of actual performance.

Glad to see you got your Iperf speeds up. I would also try and see if you can get any of your other network drivers updated. For a few more tips take a look here

If you plan on testing FTP again I might recommend you try out Filezilla. I haven't used it in a while but from what I recall it had very good performance.

Let us know how it goes.

00Roush

 

[Jeroen1000]

Although, I am getting about 100 megabit more on the office desktops, I believe that may be because they are identical on a hardware level.

I did use Filezilla for the FTP-tests. I think it was mainly a driver issue. The 'missing' 100 megabit (900 between office desktops and about 800 between my system and the laptop) can be a lot of things:

- On Vista the Ramdrive wasn't all too stable. On Win 7 it wouldn't even work. Apparently, Windows can build its own ramdrive (well surely a Microsoft driver wouldn't crash...) but I did not know that at the time.
- The network cable at work was ever so slightly better
- Driver performance is not 100% yet
- There may or may not be some driver settings to tweak
- Then we have a bunch of OS-settings we can mess about with
- Iperf is wrong

Before fiddling with any of the above, I would like to 'Iperf' between more systems. The thing is, I have only 1 gigabit capable computer in my house. So I borrowed my gf's laptop. If it is an issue with my desktop, using a different laptop won't actually help.

The good news is, at home, she has a gigabit desktop and 2 gigabit laptop's. I'll report what I find when I get the chance to hog them for a few hours:


Edit: most interesting posts you have made (talking about the tips page). I'll be reading them carefully and will try that tool you've made too.