Author Topic: Fastest tcp/ip throughput in OS/2?  (Read 3234 times)

Dariusz Piatkowski

  • Hero Member
  • *****
  • Thank You
  • -Given: 13
  • -Receive: 10
  • Posts: 596
    • View Profile
Re: Fastest tcp/ip throughput in OS/2?
« Reply #15 on: November 17, 2018, 05:02:08 pm »
Hi Lars,

At eCS times, "sack" was always set to zero. I am fairly sure it should still be this way as the feature was broken and nobody has touched AFINETK.SYS since.

And that said, that's the likely reason for the TRAP in AFINETK.SYS that you are experiencing
(make sure you use SOCKETSK.SYS and AFINETK.SYS and not SOCKETS.SYS and AFINET.SYS !)

Alright, great point. I went back to re-test this. Indeed, shutting OFF SACK prevents the TRAP from occuring with the MMRE drivers. I have not gone back to the MultiMac32-R8169 drivers yet to re-test, but will do so next.

I am pretty sure I have tried this once before, probably a good year ago when I had originaly moved from a DSL to my Fibre 125 mbps connection. At that time I found out that various sites would not connect reliably with the SACK=OFF setting. Certainly something like https://www.speedtest.net/ completely fails the test after about a 5 sec burst of traffic. Other testing sites do seem to work though.

Regardless, the normal Internet usage for me is not about using the speed testing sites, so if SACK=OFF works elsewhere it would certainly make sense to keep it that way.

Lars

  • Sr. Member
  • ****
  • Thank You
  • -Given: 0
  • -Receive: 43
  • Posts: 421
    • View Profile
Re: Fastest tcp/ip throughput in OS/2?
« Reply #16 on: November 18, 2018, 01:32:23 am »
No.
I have "Sack" disabled and www.speedtest.net works perfect  for me. You Must have a different Problem in your System.
The window Sizes should be a multiple of the MTU (or MSS,I cannot remember) and Not Just the Maximum.
« Last Edit: November 18, 2018, 01:35:55 am by Lars »

Ian Manners

  • Global Moderator
  • Sr. Member
  • *****
  • Thank You
  • -Given: 37
  • -Receive: 26
  • Posts: 321
  • I am the computer, it is me.
    • View Profile
    • ComKal Networks Australia
Re: Fastest tcp/ip throughput in OS/2?
« Reply #17 on: November 18, 2018, 07:57:57 am »
Quote
The window Sizes should be a multiple of the MTU (or MSS,I cannot remember) and Not Just the Maximum.

I use multiples of MTU, still not sure if that's right as some people say MSS, others say MTU, and the odd person says its irrelevant..
I do know if you simply use maximum values your connection can become jerky with unexpected timeouts and dropped connections.

I use manual MTU values, one day I might experiment with the window sizes to see what happens with different values and auto MTU/MSS values but it's not a high priority thing for me as what I currently have works, though an MTU of 1500 should work fine for me I've discovered the best for me is around 1000 (which ever number is closer in multiples of eight ). <== changed No. 8 to written eight
« Last Edit: November 19, 2018, 08:04:36 am by Ian Manners »
Cheers
Ian B Manners

Dariusz Piatkowski

  • Hero Member
  • *****
  • Thank You
  • -Given: 13
  • -Receive: 10
  • Posts: 596
    • View Profile
Re: Fastest tcp/ip throughput in OS/2?
« Reply #18 on: November 18, 2018, 06:43:43 pm »
No.
I have "Sack" disabled and www.speedtest.net works perfect  for me. You Must have a different Problem in your System.
The window Sizes should be a multiple of the MTU (or MSS,I cannot remember) and Not Just the Maximum.

Hmm...OK, that's promising...so what is the config you are running (inetcfg.ini)?

I am more than happy to replicate it here and report back.

BTW: Regarding the previous MMRE SACK on/off discussion...put on a couple more days on the MMRE.OS2 driver with SACK=OFF, however this still resulted in a TRAP during the upload speed test, so I'm now back to R8169.OS2, so far the same tests pass...I will put on a few more days of use and report back.

Dave Yeo

  • Hero Member
  • *****
  • Thank You
  • -Given: 11
  • -Receive: 149
  • Posts: 2156
    • View Profile
Re: Fastest tcp/ip throughput in OS/2?
« Reply #19 on: November 18, 2018, 07:03:03 pm »
I played around with speednet.net on a new install. I couldn't get it to run at all with my usual Mozilla setup, due to noscript I assume. There are a lot of scripts and some of them trigger a high CPU load. I restarted the browser in safe mode, got to see all the ads and the browser would hang after the download test with one core pegged and eventually the hung script popup would let me kill the script. Updated my hosts file which got rid of the ads but the upload test still hung. I tested both SM and FF.
Your traps might be caused by whatever was hanging the browsers here, which appears to be too many scripts with at least one our browser doesn't like.
What does work fine is the command line Python script,
Code: [Select]
wget -O speedtest-cli https://raw.githubusercontent.com/sivel/speedtest-cli/master/speedtest.py
python speedtest-cli
You may want to take the browser out of the equation or at least do tests with the Python script and compare.

Dariusz Piatkowski

  • Hero Member
  • *****
  • Thank You
  • -Given: 13
  • -Receive: 10
  • Posts: 596
    • View Profile
Re: Fastest tcp/ip throughput in OS/2?
« Reply #20 on: November 18, 2018, 07:49:25 pm »
Hey Dave!

...What does work fine is the command line Python script,
Code: [Select]
wget -O speedtest-cli https://raw.githubusercontent.com/sivel/speedtest-cli/master/speedtest.py
python speedtest-cli
You may want to take the browser out of the equation or at least do tests with the Python script and compare.

Thanks for the reminder...I did actually use this in the past, as per your earlier suggestion from a year ago, and now i went back to re-test using that process as well. I picked the same server ID to keep the results somewhat consistent with what i'm seeing across different machines on my LAN, here is what I get:

Code: [Select]
[G:\util\misc]python speedtest-cli --server 17394
Retrieving speedtest.net configuration...
Testing from Bell Canada (65.92.109.81)...
Retrieving speedtest.net server list...
Retrieving information for the selected server...
Hosted by Bell Canada (Toronto, ON) [331.11 km]: 17.667 ms
Testing download speed..........................................................
Download: 301.52 Mbit/s
Testing upload speed............................................................
Upload: 117.86 Mbit/s

The d/l values match up pretty well to what I'm seeing on other devices, the u/l however is significantly smaller. I will look into the window sizing again and see what Lars' comment might uncover.

Thanks again, this is definitely the better (more stable for sure) way of testing.

EDIT
====
Found the following testing site (http://www.dospeedtest.com/) that matches the Python script CLI test results pretty closely, so at least on my machine with FF 45.9 executing the test it appears to cause no adverse effects.

Also, I put together a quick speedtest.cmd file (see attached) that calls the Python script Dave provided and tosses-up a few entries you can chose from (runs a default test against a BELL server, which is one of the big national tel providers in Canada). You can pick your own servers of course, see the results of:

Code: [Select]
python speedtest-cli --list

...but that will be a long list, so you may want to pipe this into a file or a page-at-a-time view:

Code: [Select]
python speedtest-cli --list |less
« Last Edit: November 18, 2018, 11:57:21 pm by Dariusz Piatkowski »

Dariusz Piatkowski

  • Hero Member
  • *****
  • Thank You
  • -Given: 13
  • -Receive: 10
  • Posts: 596
    • View Profile
Re: Fastest tcp/ip throughput in OS/2?
« Reply #21 on: November 19, 2018, 02:28:16 am »
Hi Ian,

I use multiples of MTU, still not sure if that's right as some people say MSS, others say MTU, and the odd person says its irrelevant..

Yeah, as I started to look deeper into this I had the same initial reaction. But...there is good news, the inetcfg values we are looking at, specifically the 'tcprwinsize' values have their counterpart in the WIN world, that parameter is RWIN. There are a pile of articles out there that talk about the MTU <=> RWIN relationship.

Here is a link to a pretty good write-up, including a test that will check your setups and provide some walk-through analysis and recommendations, https://www.speedguide.net/analyzer.php

For now, I have throttled back my settings to as follows (just showing the relevant window size settings):

Code: [Select]
#Inetcfg: CURRENT DEFAULT MINIMUM MAXIMUM

keepalive 60 7800 0 7800 KeepAlive (sec)
tcpswinsize 127776 32768 1024 246723 TCP SendWindow Size
tcprwinsize 127776 32768 1024 246723 TCP RecvWindow Size
udpswinsize 127776 9216 1024 246723 UDP SendWindow Size
udprwinsize 127776 41600 1024 246723 UDP RecvWindow Size

Dariusz Piatkowski

  • Hero Member
  • *****
  • Thank You
  • -Given: 13
  • -Receive: 10
  • Posts: 596
    • View Profile
Re: Fastest tcp/ip throughput in OS/2?
« Reply #22 on: November 22, 2018, 04:46:47 am »
Brief update...I went back to the non-MMRE drivers, the reason being: continued TRAPs with the MMRE, even with SACK=OFF.

As Lars pointed out, could be my overall settings (althought I backed down on the window sizing, using about 1/2 of the buffer now as compared to where I started), or it could be that the generic Unix to OS/2 layer that allows the bigger drivers pool to be utilized has some inherent problems that I'm seeing symptoms of.

Anyways, just FYI re: my experience, if anyone else has a well working high-speed setup please post your config.

guzzi

  • Sr. Member
  • ****
  • Thank You
  • -Given: 0
  • -Receive: 18
  • Posts: 303
    • View Profile
Re: Fastest tcp/ip throughput in OS/2?
« Reply #23 on: November 23, 2018, 01:56:16 am »
My connection is not nearly as fast, but I get the max speed it allows:
Testing download speed..
......................
Download: 60.96 Mbit/s
Testing upload speed....
........................
Upload: 19.17 Mbit/s