This is the mail archive of the cygwin mailing list for the Cygwin project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Cygwin TCP slow


On 2016-11-29 19:38, Sam Habiel wrote:
> On Tue, Nov 29, 2016 at 8:08 PM, Daniel Havey <dhavey@gmail.com> wrote:
>> On Mon, Nov 28, 2016 at 12:51 PM, Brian Inglis
>> <Brian.Inglis@systematicsw.ab.ca> wrote:
>>> On 2016-11-28 12:54, Daniel Havey wrote:
>>>> We have had complaints from several large hardware vendors
>>>> that Windows networking is slow for apps like iperf that are
>>>> used to measure throughput. Iperf on Windows is compiled
>>>> against the cygwin1.dll. We have root caused the problem to a
>>>> couple of lines of code in net.cc that set SO_RCVBUF and
>>>> SO_SNDBUF to about 200KB.
>>>>
>>>> The theoretical window/RTT plot for the buffer size set by Cygwin
>>>> (0x34000 = 200KB) gives us:
>>>> 1ms -> 1703Mbps
>>>> 2ms -> 851Mbps
>>>> 3ms -> 567Mbps
>>>> 4ms -> 425Mbps
>>>> 5ms -> 340Mbps
>>>> 6ms -> 283Mbps
>>>> 7ms -> 243Mbps
>>>> 8ms -> 212Mbps
>>>> 9ms -> 189Mbps
>>>> 10ms -> 170Mbps
>>>> 20ms -> 85Mbps
>>>> 40ms -> 42Mbps
>>>> 60ms -> 28Mbps
>>>> 80ms -> 21Mbps
>>>>
>>>> We have confirmed this by experiment and also confirmed that
>>>> the limitation goes away if the buffers are not manually set.
>>>> Windows has autotuning and when the buffers are set manually
>>>> the autotuning is disabled. This is causing the throughput
>>>> limitation. So we would like to formally ask that you please
>>>> not manually set SO_RCVBUF or SO_SNDBUF.
>>>
>>> See problem reports: http://cygwin.com/problems.html
>>> Provide STC, patches, attach cygcheck -svr output?
>>> Links to downstream bug reports, testing, results?
>>> Note that Cygwin iperf is year old 2.0.5.

>> Okay, I will find some time to produce a patch. It might take a
>> while though because I have a day job :). BTW, what the heck is an
>> STC?

STC == Simple Test Case

>> Here is an experiment with three machines like this: O----O----O
>> The one in the middle has a 50ms of delay (25ms in each direction).
>>
>> Here are the results from Cygwin on top of normal Windows:
>> [send side perf]
>> f:\home>iperf3 -c 10.178.204.101
>> [ ID] Interval           Transfer     Bandwidth
>> [  4]   0.00-10.00  sec  47.5 MBytes  39.8 Mbits/sec                  sender
>> [  4]   0.00-10.00  sec  47.5 MBytes  39.8 Mbits/sec                  receiver
>>
>> [receive side perf]
>> f:\home>iperf3 -c 10.178.204.101 -R
>> [ ID] Interval           Transfer     Bandwidth       Retr
>> [  4]   0.00-10.00  sec  41.2 MBytes  34.5 Mbits/sec    0             sender
>> [  4]   0.00-10.00  sec  39.7 MBytes  33.3 Mbits/sec                  receiver
>>
>> This matches the calculated performance. Then we made a private 
>> build of Windows that ignores SO_RCVBUF and SO_SNDBUF and just 
>> always uses autotuning no matter what the app does.

> Out of curiosity, how did you do that? Do you work for Microsoft, or 
> is there something fantastic I missed about building your own 
> modified DLLs?

Or is this an Enterprise or Insider build with specific GPOs or 
reg entries that could be applied on any Windows release?

You state the basics nowhere - Windows release and flavour, 
Cygwin release and arch, installed packages, DLLs, et al.
That's the reason for the problem reports web page asking for 
cygcheck -svr output to be attached, and the lack of any 
attachment may mean your post may be ignored by the tiny
group of very busy Cygwin volunteers.

If the problem is on XP, Cygwin no longer supports it, and if 
it only occurs on Vista or 8, and not reproducible on 7 or 10, 
it may be weighed accordingly.
If it only affects artificial benchmarks like iperf3 and not 
real network apps like say rsync, netcat, etc. it may be not be 
considered high enough impact to look at, among higher priorities.

If you could reproduce this with an available package like rsync 
and equivalent settings, it may get higher priority.

If you also provided a Cygwin port of iperf3 you could win a 
*Gold Star* from the Cygwin volunteers who may then be able 
to reproduce your results.

>> [send side perf]
>> C:\testbox\tests>iperf3 -c 10.178.204.101
>> [ ID] Interval           Transfer     Bandwidth
>> [  4]   0.00-10.00  sec   540 MBytes   453 Mbits/sec                  sender
>> [  4]   0.00-10.00  sec   540 MBytes   453 Mbits/sec                  receiver
>>
>> [receive side perf]
>> C:\testbox\tests>iperf3 -c 10.178.204.101 -R
>> [ ID] Interval           Transfer     Bandwidth       Retr
>> [  4]   0.00-10.00  sec   553 MBytes   464 Mbits/sec    0             sender
>> [  4]   0.00-10.00  sec   553 MBytes   464 Mbits/sec                  receiver
>>
>> If you set SO_RCVBUF and SO_SNDBUF then performance will be
>> limited according to the calculated values. If you don't set those
>> values and let Windows autotuning do its thing then you will always
>> get the maximum available throughput.
>>
>> I'll email again when I have the patch. If you would like more 
>> testing let me know and we can have our test people run some more 
>> experiments.

If you attach cygcheck -svr output, and can reproduce in a STC with an 
available widely used Cygwin package, you are more likely to get read 
by one of the Cygwin committers or volunteers who can help.

-- 
Take care. Thanks, Brian Inglis, Calgary, Alberta, Canada

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]