Bug 789659

Summary: limited network bandwidth
Product: [Fedora] Fedora Reporter: nucleo <alekcejk>
Component: kernelAssignee: Kernel Maintainer List <kernel-maint>
Status: CLOSED ERRATA QA Contact: Fedora Extras Quality Assurance <extras-qa>
Severity: unspecified Docs Contact:
Priority: unspecified    
Version: 15CC: gansalmon, itamar, jonathan, kernel-maint, madhu.chinakonda
Target Milestone: ---   
Target Release: ---   
Hardware: Unspecified   
OS: Unspecified   
Whiteboard:
Fixed In Version: Doc Type: Bug Fix
Doc Text:
Story Points: ---
Clone Of: Environment:
Last Closed: 2012-05-07 20:51:28 UTC Type: ---
Regression: --- Mount Type: ---
Documentation: --- CRM:
Verified Versions: Category: ---
oVirt Team: --- RHEL 7.3 requirements from Atomic Host:
Cloudforms Team: --- Target Upstream Version:
Embargoed:

Description nucleo 2012-02-12 04:20:29 UTC 
Description of problem:
After updating to kernel-2.6.42.3-2.fc15.i686 download speed from some internet resources became about half of speed with kernel-2.6.41.10-3.fc15.i686.

Version-Release number of selected component (if applicable):
kernel-2.6.42.3-2.fc15.i686

How reproducible:
sometimes

Actual results:
Download speed of ftp://ftp3.de.freebsd.org/pub/FreeBSD/ISO-IMAGES-i386/8.2/FreeBSD-8.2-RELEASE-i386-disc1.iso:
with kernel-2.6.42.3-2.fc15.i686 - 5.82 MB/s (speed almost constant at download time)
with kernel-2.6.41.10-3.fc15.i686 - average 8.74 MB/s (average), at maximum 10.7 MB/s

Download speed of ftp://ftp.snt.utwente.nl/pub/os/linux/ubuntu-releases/11.10/ubuntu-11.10-alternate-amd64.iso:
with kernel-2.6.42.3-2.fc15.i686 - 5.70 MB/s (constant)
with kernel-2.6.41.10-3.fc15.i686 - 7.99 MB/s (average), about 11 MB/s at maximum

Download speed of http://se.releases.ubuntu.com/11.10/ubuntu-11.10-alternate-i386.iso a bit more with 2.6.42.3:
with kernel-2.6.42.3-2.fc15.i686 - 6.70 MB/s (constant)
with kernel-2.6.41.10-3.fc15.i686 - 8.06 MB/s (average), about 11 MB/s at maximum

But from some addresses sped with 2.6.42.3 almost the same as with 2.6.41.10.
Download speed from http://ftp.linux.kiev.ua/pub/Linux/Ubuntu/releases/11.10/ubuntu-11.10-alternate-amd64.iso:
with kernel-2.6.42.3-2.fc15.i686 - 11.2 MB/s (constant)
with kernel-2.6.41.10-3.fc15.i686 - 11.3 MB/s (constant)


Expected results:
Download speed with new kernel should not be less than with old.
Additional info:

lsmod:
Module                  Size  Used by
nvidia              10778232  40
i2c_core               28123  1 nvidia
8139too                32490  0
mii                    13311  1 8139too

dmesg:
8139too: 8139too Fast Ethernet driver 0.9.28
8139too 0000:01:02.0: PCI INT A -> GSI 22 (level, low) -> IRQ 22
8139too 0000:01:02.0: eth0: RealTek RTL8139 at , IRQ 22
8139too 0000:01:02.0: eth0: link up, 100Mbps, full-duplex, lpa 0x41E1

I rebooted many times with old kernel and new and every time download speed from some addresses with 2.6.42.3-2 was lower than with 2.6.41.10-3.
Don't know is this related with 8139too driver or maybe with some protocols drivers.
Comment 1 Dave Jones 2012-02-13 16:19:11 UTC 
the changes to 8139too between these kernels are fairly trivial, and wouldn't explain any kind of performance difference.

I get different throughput each time I try to connect to those sites (varying from 1.3M/s to 2M/s). Given there may be load balancing happening on their side, I don't think there's a clear sign that this is a bug.
Comment 2 nucleo 2012-02-13 16:34:50 UTC 
Here site that I also tested with both kernels: http://speedtest.tele2.net/ which have enough network bandwidth and good enough hardware/software for speed tests.
With 2.6.41.10-3 I got there maximum speed 11 MB/s but with 2.6.42.3-2 speed again limited with 5.5 MB/s.
And again I want to pay attention that limited speed stays constant - always 5.5 MB/s, straight line on traffic plotter as if was network bandwidth shaping. With 3.1 kernels this never happens.
So there is definitely some kind of regression in 3.2 kernels.
Comment 3 Dave Jones 2012-02-13 16:45:56 UTC 
ok, I recommend taking this to netdev.org.
For reference, let them know 2.6.41 = 3.1, and 2.6.42 = 3.2
Comment 4 nucleo 2012-02-13 17:47:49 UTC 
You mean that I should send mail to netdev.org or subscription required there at first?
Comment 5 Dave Jones 2012-02-13 19:11:52 UTC 
no subscription necessary, it's an open list.
Comment 6 nucleo 2012-03-23 21:55:37 UTC 
Problem still in kernel-3.3.0-1.fc17.i686
Comment 7 Dave Jones 2012-05-07 18:50:18 UTC 
Did this get solved in 3.4 ? I saw your posts to netdev, but don't recall the resolution.
Comment 8 nucleo 2012-05-07 19:05:46 UTC 
This was solved in kernel 3.3.4 by changing net.ipv4.tcp_rmem value for 32bit kernel from "4096 87380 897664" to "4096 87380 4194304" that already was used for 64bit kernel. This solved speed limitation for large RTT.

But in my tests with 3.1 kernel net.ipv4.tcp_rmem="4096 87380 897664" looks better than net.ipv4.tcp_rmem="4096 87380 4194304" because with 897664 I got constant maximal speed but with 4194304 speed always changed from maximal to lower values, so average speed was less than with 897664.
Comment 9 Dave Jones 2012-05-07 20:51:28 UTC 
ok. if there are further changes needed, let the netdev guys know and we'll pick them up as they get merged.

thanks for your work on this.