[Snap] SNAP performance to the USA?

Forums › 2degrees (including Slingshot, Orcon, Flip, Stuff Fibre, MyRepublic, 2talk and Vocus) › [Snap] SNAP performance to the USA?

1 | 2

sidefx

3712 posts

Uber Geek

Trusted

#674190 19-Aug-2012 11:30

Interesting... I alternated speedtests between a windows 7 box and ubuntu box both last night and this morning and seemed to get fairly consistently higher results on linux. My windows 7 machine has a lot more grunt that linux machine.

WINDOWS:

LINUX:

"I was born not knowing and have had only a little time to change that here and there." | Octopus Energy | Sharesies
- Richard Feynman

mercutio

1392 posts

Uber Geek

#674193 19-Aug-2012 11:36

spacedog:
Ragnor:
spacedog:?

Can't see how operating system would have any effect. I can pull 14mbps on a speedtest to an auckland server on the same computer and then run a speedtest to a USA server and only get 4-5mbps. ?If the OS was a factor, I wouldn't be getting the fast domestic speeds.

The TCP/IP stack has been changed and improved with newer versions of windows, most notably with automatic tcp receive window scaling.

Refer to?
http://www.speedguide.net/articles/the-tcp-window-latency-and-the-bandwidth-delay-2678?

http://www.speedguide.net/articles/windows-7-vista-2008-tweaks-2574

So this most likely means that difference I am seeing in side-by-side speedtesting of domestic/international traffic from a windows7 machine is not being affected by the OS, right??

As a test, I ran some bandwidth testing from two different linux machines on my home network and didn't see any speeds over 6mbps to international routes....

Linux autotunes based upon memory, which mostly just undertunes if you have faster server on a tiny VPS or router with hardly any RAM.

Download speeds are based upon /proc/sys/net/ipv4/tcp_rmem which you can also set in /etc/sysctl.conf.

Linux 3.0 also started setting the initial send window size larger, which should make short uploads slightly faster. Some versions past then added more stuff.

mercutio

1392 posts

Uber Geek

#674194 19-Aug-2012 11:39

sidefx: Interesting... ? I alternated speedtests between a windows 7 box and ubuntu box both last night and this morning?and seemed to get fairly consistently higher results on linux. Hmm.. maybe an issue with the speedtest\flash on linux? ?Or my antivirus on windows? ?Or some configuration somewhere? ? My windows 7 machine has a lot more grunt that linux machine.

it's not really very cpu hungry unless you're doing gigabit speeds any CPU should do.

Could be configuration. I'd try setting the window size autoscaling slightly higher first. And enable ctcp.

At 180 msec it could just be that Linux is recovering from packet loss faster though.

11:45:11.747276 216.110.200.226.80 > : S 638669006:638669006(0) ack 1093069853 win 14480 (DF)

With a tcpdump you can see that that OR server advertises a default window size of 14480 bytes - which is 10xmss with mss of 1460. The window size seems slightly lower than I'd imagine 14600 would seem normal - but that's the same amount of overhead as tcp/ip timestamps.. so maybe it's remembering support for that.

Anyway, it's normal for at least Linux and OpenBSD on recent kernels to advertise the larger window size. And Windows has had a larger window size for ages. This'll effect download/upload speeds for short transfers the most.

But that could be one of the reasons that server gives good performance.

sidefx

3712 posts

Uber Geek

Trusted

#674200 19-Aug-2012 11:48

mercutio:
Could be configuration. I'd try setting the window size autoscaling slightly higher first. And enable ctcp.

I have ctcp enabled with not noticable effect (and most other settings in Ragnors 2nd link set to the "recommended" values)

How would I go about setting window size autoscaling "slightly higher"?

"I was born not knowing and have had only a little time to change that here and there." | Octopus Energy | Sharesies
- Richard Feynman

mercutio

1392 posts

Uber Geek

#674206 19-Aug-2012 11:53

sidefx:
mercutio:?
Could be configuration. I'd try setting the window size autoscaling slightly higher first. And enable ctcp.

I have ctcp enabled with not noticable effect (and most other settings in Ragnors 2nd link set to the "recommended" values) ?

How would I go about setting window size autoscaling "slightly higher"?

Oh it only has normal and expiremental! maybe i was wrong...

Just as a check that you have no local problems on your windows pc, I'd do a local iperf speed test in each direction between windows/linux.

just checking tcpdump on my windows 8 host, it seems window size is getting up to 64k pretty much. that could be happening to you too and explain the slow speeds.

actually windows 8 is increasing to 1mb/sec window size. and getting packet loss, maybe it's best to actually set window size smalelr...

linux is getting up to window size of 930k... so that's similar... i have tcpdump outputs for both, still confused why linux speeds up WAY faster... and ends up a few hundred k/sec faster.

1 | 2