Need help with understanding assymetric performance figures

[flyvert]

Hi,

My Ultra2 is now about a year old - the performance hasn't really been what I expected from start, but I have not had time to check it out properly until tonight.

I'm running 4.2.22 and have two of those WD Greens 2TB (but I've managed to get them stop parking their heads all the time using the idlectl tool I found on this forum) which I know are not on the certified hardware list.

Drag & drop from Windows 7 to the NAS over a Wireless-N net typically gives write performance about 17-20 MB/s.

I have now tested using ReadNAS's iometer.icf file and IOMeter via:
a) WLAN (802.11n, 5GHz band)
b) wired (gigabit LAN through my router/switch)
c) wired straight connection

I'm not sure how to intepret the results...

Computer Net      Jumbo Read Write 
Laptop   VLAN     No    14   18
Laptop   LAN      Yes   24   28
Laptop   LAN      No    24   29
Laptop   Straight Yes   35   25
Desktop  LAN      No    41   41
Desktop  LAN      Yes    0    0
Desktop  Straight No    44   15
Desktop  Straight Yes    0   50

My stationary supports jumbo frames starting with 2K up to 7K; I tested 2, 5, 7 with very similar results.
The NAS was at 9K when jumbo frames was enabled.

Oddities:

- Jumbo frames did not work well via LAN and my stationary

- Enabling jumbo frames in both ends and using a straight connection (my laptop has Auto-MDIX) gives 50 MB/s write performance, but zero (0) read!

- My 2008 high end laptop's NIC gives less performance than my very low end stationary's NIC (both are Gigabit)


Can someone please try to explain the strange results?

Are these expected values for an Ultra2?

/f

[StephenB]

For large files it should be possible to get 70 MB/s read and 50 MB/s write. Multiple small files are a lot slower. Though the performance of your laptop/PC can be the limiting factor (either hard drive speed, security/antivirus software, or network).

I have no explanation for the failure of jumbo frames on PC reads - though you might want to troubleshoot it with transferring data between the PC and the Laptop. You should also try confirming the MTU with ping (especially when connected to the router).

You also might look at the advanced NIC settings, and try turning off the "offload" features. Sometimes they help throughput, other times not so much.

You might also try an alternative to iometer. Teracopy for instance can time transfer speed.

[HERBIEO]

LAN Speed Test is a pretty good for testing performance they do LAN Speed Test (Lite) (Free version) i purchased the full version with the LST Server Add on which is very usefull for testing the real throughput of your router without the bottleneck of Hard disks involved take a look here > http://www.totusoft.com/index.html

[etardieu]

As written in a post above:
ultra 2, old (2009) samsung 1to disks in RAID1, Gigabit eth with 9k jumbo frames:
100Mo/s read CIFS
60Mo/s write CIFS
client is windows 8 with SSD. You are definitely not getting expected performance.
emmanuel

[mdgm-ntgr]

The Ultra 2 was released in 2010 not 2009. You couldn't have got one in 2009.

What disk is in your PC?

[etardieu]

mdgm wrote:
The Ultra 2 was released in 2010 not 2009. You couldn't have got one in 2009. What disk is in your PC?

the disks are from 2009. My ultra 2 is brand new.

[mdgm-ntgr]

Ah oops I misread your post. I see that now.

What disk is in your Windows PC?

Are your PC NIC drivers up to date?

[StephenB]

I didn't think etardieu was complaining about performance??? 100 MB/s reads and 60 MB/s writes for an ultra-2 seems pretty good to me.

flyvert's setup seems to be under-performing though.

[etardieu]

nope, not complaining! I was giving some perofrmance feedback to the OP. Sorry if I didn't make it clear.

[flyvert]

As a house owner with small kids I seldomly get time to dwelve into this, but since IOmeter got 50 MB/s in write performance (but 0 read!) when using a straight cable and jumbo frames there seemes to be some "juice" in the box?

My guess is NIC HW/SW settings or incompatibilites in the jumbo feature implementation. The NAS do 9k but my desktop's low end NIC only 7k (desktop and laptop disks are both mid range Intel SSD so they should not have tarnished the figures in my first post to this thread)

LAN figures involves a NETGEAR GS108 8-port gigabit switch, WLAN a Cisco Linksys E4200 router - at this stage I don't expect them to be pulling down the figures (definitely not when using a straight cable...)

At next opportunity (kids in their beds and spouse out of the house) I will do some more testing with other perf tools, wires and NICs...

I have two years left on the NAS's HW warranty and I also notised that the store I bought NAS and disks at now carry the WD reds (too bad they didn't have them in stock when I got my greens replaced due to excessive head parking some months ago).

If I can't get these performance figures to rise I will probably select to get the whole box replaced since I also have an issue with one of the disk indicator LEDs that is broken). However, the store estimated it may take 2-3 weeks during which I can't play music with my squeeze box, etc.

/f

[flyvert]

Yet some more tests performed and I'm not sure how to interpret the results...

This time I decided to use the "normal setup" for all measurements:
- Desktop P4 3.4 GHz anno 2003 (yeah, its crap - but fulfills my requirements, so far...)
- Windows 7 SP1 w all updates,
- Microsoft Security Essentials disabled
- Intel SSD disk benchmarked to ~100 MB/s with HDTune (I only got a cheap SATA1 card to interface it with)
- Realtek PCI GBE with latest driver

The 10 meter LAN is first passing a Cisco Linksys E4200 switch and then also a NETGEAR GS108 switch before reaching the Ultra 2 NAS.

LAN Speed Test (Lite) version 1.2.0
(Jumbo frames disabled)
Write: 45 MB/s
Read: 49 MB/s

(Jumbo frames enabled)
Write: 6 MB/s
Read: x (failed!)

Teracopy
(Jumbo frames disabled)
Write: 22 MB/s
Read: 7 MB/s

Windows 7 copy using Explorer
(Jumbo frames disabled)
Write: 15 MB/s
Read: 8 MB/s

When using Windows 7 copy, the CPU is almost saturated to 100% while LAN Speed Test is using <30%.
The biggest CPU hogs are "explorer.exe ~50%, taskmgr ~15% and dwm (window manager) ~10%

The network tab of Task Manager shows two reasonable looking "squares" when using LAN Speed Test:



When copying the same file using Windows Explorer it looks like this:



Notice the change in Y-scale and behavior of "up" vs "down" squares.

What conclusions may a draw from this?

I guess the LAN Speed Test does not have to worry about reading/writing any local file during the test, only the remote.

- Why is LAN Speed Test so much faster at uploading a gigabyte compared to the other two?

- Jumbo Frames does not seem usable in my network

What can I do to improve "daily usage", i.e. reading/writing files from Windows 7 Explorer?

/f

[HERBIEO]

LAN Speed Test does this by building a file in memory, then transfers it both ways without effects of windows file caching so you can see the real transfer speed across the network if you use the LST server addon you can install that on another computer on your network and run it then you get the effect both ways so you can see the real throughput of your router and switches, what you have to remember is the slowest bottleneck is most likely having to read and write the data to and from hard disks so this takes the hard disks out of the equation, i found it to be a very usefull little program thats why i purchased the package of the full version with the LST server addon only $11.00

[flyvert]

Thanks Herbieo,

I'd like to know if anything could be done to improve the "daily transfers" made using Mr Ballmer's regular desktop tools?

Since other tools may get thru as much as 50 MB/s the bottleneck seems to be in the way Windows File copying works.

My old old stationary's 15 / 8 and my high end laptop's 29 / 24 seen when "Ballmer copying" is only 50% of what the performance tools are hinting "is there" plus the odd fact that writing is faster than reading!

Anyone with a Realtek PCI GBE adapter in an older Ballmer box having more luck than I?

/f

[StephenB]

You already know one problem - your older windows machine is running out of CPU.

The NIC card should have a variety of "offload" CPU options you can try setting. (interrupt moderation, checksum stuff) You might try turning all of them on and see if that helps.

Also, how much memory is in desktop PC?

[flyvert]

StephenB wrote:
You already know one problem - your older windows machine is running out of CPU. The NIC card should have a variety of "offload" CPU options you can try setting. (interrupt moderation, checksum stuff) You might try turning all of them on and see if that helps.

Thanks for the feedback Stephen,

Shaved off every thing (except for fiddling with interrupts) from the Realtek NIC last night to no avail.
Got even less thru with the offloading and large file transfer settings disabled.

Even if my P4 "belongs in a museum" my high end laptop is also 50% away from what some perf tools says is there (even when using my old stationary box).

I wonder if the methods used by these tools even are comparable with Windows File copying? File2File is not the same as C#2File.

I also wonder why jumbo frames are making things worse for me when the broader audience seem happy with it?

/looking at some motherboard/processor/RAM deals online

[StephenB]

flyvert wrote:
Shaved off every thing (except for fiddling with interrupts) from the Realtek NIC last night to no avail. Got even less thru with the offloading and large file transfer settings disabled.

In my experience that is often the result - they reduce CPU load, but they often decrease throughput. I was hoping with your older P4 that you might get a better result.

flyvert wrote:
I wonder if the methods used by these tools even are comparable with Windows File copying? File2File is not the same as C#2File.

Good question. Also the file size (disk caching) can make a huge difference. It would be really nice if there was a true network test built into the ReadyNAS (perhaps via an add-on) and an associated PC client, so you could separate out the raw LAN performance from the rest of the puzzle. Ideally there would be a layered set of tests that would make this troubleshooting much more systematic.

flyvert wrote:
I also wonder why jumbo frames are making things worse for me when the broader audience seem happy with it?

You are not alone, there are lots of posts here from folks who have found that jumbo frames make things worse. Though your issue with writes working and reads failing does seem strange. Anyway, I think something along the path is fragmenting the jumbo frames. Have you tried probing the MTU with ping?