[SOLVED]Slow speeds between two Connectx-2 machines

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

T_Minus

Build. Break. Fix. Repeat
Feb 15, 2015
7,640
2,057
113
I just replaced the lamps and cable with OM3 cable as suggested above by @Quasduco and I'm getting 8+Gbps now via iperf between the two machines, and 300-400 MB/s over SMB to my fileserver. Now, I'm a happy camper. Thanks to everyone for their help :)
Awesome!

Which cables were you using before and if you recall which seller? Might help me, and someone else out in the future :) :)
 

rubylaser

Active Member
Jan 4, 2013
846
236
43
Michigan, USA
Awesome!

Which cables were you using before and if you recall which seller? Might help me, and someone else out in the future :) :)
I used this cable originally. It shipped fast and was well packaged, it just didn't work for me.

BELKIN 10M SFP+10GBASE DirectAttachPassive Twinaxial Cable from hypermicrosystems
 
  • Like
Reactions: T_Minus

epicurean

Active Member
Sep 29, 2014
785
80
28
Hi Rubylaser,
May I know exactly which adapters you used and the OM3 cable that came together for you?
 

William

Well-Known Member
May 7, 2015
789
252
63
66
So bumping an old post.
Seems just about all google searchs about ConnectX-2 slow speeds point back to STH, there are a number of posts about this here.

I have been battling this problem for a long time now with no real effects to get me to 10G speeds. I picked up a ConnectX-3 Pro the other day to give it a tray (thanks Patrick for info on the SFP+ adapter).

System is like this.
DS1618+ with ConnectX-3 (this came with the unit)
Workstation - ConnectX-3 Pro
Switch D-Link DXS-1210-12TC
Mellanox drivers 5.50 (I have tried earlier versions also)

I am lucky to get 2G speeds, even if I direct connect workstation to NAS.

A year ago or so I picked up a pair of ConnectX-2 single port cards that came with cables which are CISCO SFP-H10GB-CU2M.

So after all this testing and trying again when I got the ConnectX-3 Pro card I came across this thread.

Is it possible that these NIC's just don't like the cables I have ?
Is it worth while to grab a set of these ?
Cisco SFP+ Copper Twinax Cable - Twinaxial 5 m for Catalyst 2960, 2960-24, 2960-48, 2960G-24, 2960G-48, 2960S-24, 2960S-48 | eBay

Or maybe some settings on my Workstation need to be changed, motherboard is a Z10PE-D16 WS, OS Win 10 Pro.
The card is in the first slot which is a PCIe Gen3 x8 and SR-IOV is enabled in the BIOS.
All LEDs on the NIC's and Switch show green (10G).

iperf results new.JPG
 

i386

Well-Known Member
Mar 18, 2016
4,241
1,546
113
34
Germany
I have been battling this problem for a long time now with no real effects to get me to 10G speeds.
With actual workloads? Or iperf?

On windows I never really got good results in iperf on the windows machines, but had no problems to get ~ 9gbit/s transfers between workstations with iodrive2s.
 

arglebargle

H̸̖̅ȩ̸̐l̷̦͋l̴̰̈ỏ̶̱ ̸̢͋W̵͖̌ò̴͚r̴͇̀l̵̼͗d̷͕̈
Jul 15, 2018
657
245
43
Those numbers look about right for a single stream iperf test from windows. I think I got something close to that before I switched to using 8 parallel streams, in parallel I got very close to full line speed.
 
Last edited:

William

Well-Known Member
May 7, 2015
789
252
63
66
Workloads are nothing special. Something like a DVD ISO copy from the NAS to workstation takes like a minute, under 10G it should be like 10 seconds.

Let me look at how many threads are being used.
 

William

Well-Known Member
May 7, 2015
789
252
63
66
Max number of RSS Processors is set to 8
RSS Base Processor Number is set to 8

I think those are the right ones to change ?

This showed no difference is speed. I will restart when I can and see if that fixes it.
 

arglebargle

H̸̖̅ȩ̸̐l̷̦͋l̴̰̈ỏ̶̱ ̸̢͋W̵͖̌ò̴͚r̴͇̀l̵̼͗d̷͕̈
Jul 15, 2018
657
245
43
Also, with jumbo frames I lost a significant amount of bandwidth on a direct connection between two cards, I think I dropped from 39Gb/s to like 14Gb/s with jumbo frames on. If you're using 9k MTU test with 8 stream iperf and test 1500 MTU too.
 
Last edited:

arglebargle

H̸̖̅ȩ̸̐l̷̦͋l̴̰̈ỏ̶̱ ̸̢͋W̵͖̌ò̴͚r̴͇̀l̵̼͗d̷͕̈
Jul 15, 2018
657
245
43
Max number of RSS Processors is set to 8
RSS Base Processor Number is set to 8

I think those are the right ones to change ?

This showed no difference is speed. I will restart when I can and see if that fixes it.
You need to use -P 8 on the iperf client command line. That's it, don't mess with nic RSS settings.
 

arglebargle

H̸̖̅ȩ̸̐l̷̦͋l̴̰̈ỏ̶̱ ̸̢͋W̵͖̌ò̴͚r̴͇̀l̵̼͗d̷͕̈
Jul 15, 2018
657
245
43
I used this
iperf3.exe -w 1024k -P 8 -i 2 -t 30 -c 10.0.0.198
Yup, that should do it. I don't remember if you need to run the server with -P 8 too, I think I had to for one of the iperfs (iperf or iperf3) but I can't recall which.

I didn't need to specify window size, I just let it rip with the defaults.
 

saivert

Member
Nov 2, 2015
138
18
18
40
Norway
There is no need to run multithreaded iperf tests. ConnectX-2 and ConnectX-3 is capable of maxing at 9.8Gbps single stream. Just make sure your PC is up to the task at it takes some CPU power. I had an issue with my old PC with Core i7 2600k OC-ed to 4.3GHz where just opening up MPC-HC (video player) made the test drop to 4.8Gbps for RX test. Do make sure you have recent firmware on them. I didn't get full speed on my ConnectX-3 until it was upgraded.
 

arglebargle

H̸̖̅ȩ̸̐l̷̦͋l̴̰̈ỏ̶̱ ̸̢͋W̵͖̌ò̴͚r̴͇̀l̵̼͗d̷͕̈
Jul 15, 2018
657
245
43
There is no need to run multithreaded iperf tests. ConnectX-2 and ConnectX-3 is capable of maxing at 9.8Gbps single stream. Just make sure your PC is up to the task at it takes some CPU power. I had an issue with my old PC with Core i7 2600k OC-ed to 4.3GHz where just opening up MPC-HC (video player) made the test drop to 4.8Gbps for RX test. Do make sure you have recent firmware on them. I didn't get full speed on my ConnectX-3 until it was upgraded.
Here's iperf3 with 8 streams (2 threads according to top):
iperf 40gb 5m SAS.png iperf 40gb 2m SAS.png

Here's iperf3 non-parallel:
iperf3 CX3 netapp SAS 5m mtu 1500.png iperf3 CX3 netapp SAS 2m mtu 1500.png

The CPU on the receiving linux machine bottlenecked the 8 stream transfer to ~35Gb/s, that's the maximum I was able to get out of that point to point link. 8 streams gets the job done without needing ludicrous single thread performance to do it.

The server is a G4400 Pentium, pushing 35Gb/s.
 
Last edited:

William

Well-Known Member
May 7, 2015
789
252
63
66
WOW very nice.

Like I said above, my system is Z10PE-D16 WS, OS Win 10 Pro. CPU's are 2x E5-2699v3 and 256GB RAM installed.
Although these are not fast CPU's there is a lot of cores/threads.