[OpenAFS-devel] .35 sec rx delay bug?

Sven Oehme oehmes@de.ibm.com
Fri, 3 Nov 2006 20:56:30 +0100


This is a multipart message in MIME format.
--=_alternative 006D449BC125721B_=
Content-Type: text/plain; charset="US-ASCII"

i mentioned that a few month ago, if you configure the network, client and 
server right and you have the right HW in place you can get to ~ 50 
Mbyte/sec from Client to Server over Gigabit. and btw. i saw that delay 
bug too and i can confirm that it has something to do with the HW you use, 
if that helps i only saw it on fast SMP boxes ..

Sven




Ken Hornstein <kenh@cmf.nrl.navy.mil> 
Sent by: openafs-devel-admin@openafs.org
11/03/06 08:06 PM

To
Jim Rees <rees@umich.edu>
cc
openafs-devel@openafs.org
Subject
Re: [OpenAFS-devel] .35 sec rx delay bug?






>  350ms is the minimum retransmit delay. Is it possible that your network 
is 
>  dropping packets?
>
>It's possible, and I'll check.  But a 10MB file takes 2 seconds to 
transfer
>using ssh over tcp, and over 4 minutes using afs over rx over udp.  Is rx
>really that much worse than tcp?

I saw a paper that Sine Nomine did for a customer on AFS client
performance (the work is in the public domain, but it's actually rather
hard to find a copy of this paper).  One thing they discovered is that
when network performance went above 100 megabits, Rx throughput fell off
erratically (on Gig-E, they observed an average throughput of 15 Mb/s).
So yes ... rx really is that much worse than TCP.  I suspect that this
is a long-standing bug that has only manifested itself recently with the
advent of faster disks and network.

--Ken
_______________________________________________
OpenAFS-devel mailing list
OpenAFS-devel@openafs.org
https://lists.openafs.org/mailman/listinfo/openafs-devel


--=_alternative 006D449BC125721B_=
Content-Type: text/html; charset="US-ASCII"


<br><font size=2 face="sans-serif">i mentioned that a few month ago, if
you configure the network, client and server right and you have the right
HW in place you can get to ~ 50 Mbyte/sec from Client to Server over Gigabit.
and btw. i saw that delay bug too and i can confirm that it has something
to do with the HW you use, if that helps i only saw it on fast SMP boxes
..</font>
<br>
<br><font size=2 face="sans-serif">Sven<br>
</font>
<br>
<br>
<br>
<table width=100%>
<tr valign=top>
<td width=40%><font size=1 face="sans-serif"><b>Ken Hornstein &lt;kenh@cmf.nrl.navy.mil&gt;</b>
</font>
<br><font size=1 face="sans-serif">Sent by: openafs-devel-admin@openafs.org</font>
<p><font size=1 face="sans-serif">11/03/06 08:06 PM</font>
<td width=59%>
<table width=100%>
<tr valign=top>
<td>
<div align=right><font size=1 face="sans-serif">To</font></div>
<td><font size=1 face="sans-serif">Jim Rees &lt;rees@umich.edu&gt;</font>
<tr valign=top>
<td>
<div align=right><font size=1 face="sans-serif">cc</font></div>
<td><font size=1 face="sans-serif">openafs-devel@openafs.org</font>
<tr valign=top>
<td>
<div align=right><font size=1 face="sans-serif">Subject</font></div>
<td><font size=1 face="sans-serif">Re: [OpenAFS-devel] .35 sec rx delay
bug?</font></table>
<br>
<table>
<tr valign=top>
<td>
<td></table>
<br></table>
<br>
<br>
<br><tt><font size=2>&gt; &nbsp;350ms is the minimum retransmit delay.
Is it possible that your network is <br>
&gt; &nbsp;dropping packets?<br>
&gt;<br>
&gt;It's possible, and I'll check. &nbsp;But a 10MB file takes 2 seconds
to transfer<br>
&gt;using ssh over tcp, and over 4 minutes using afs over rx over udp.
&nbsp;Is rx<br>
&gt;really that much worse than tcp?<br>
<br>
I saw a paper that Sine Nomine did for a customer on AFS client<br>
performance (the work is in the public domain, but it's actually rather<br>
hard to find a copy of this paper). &nbsp;One thing they discovered is
that<br>
when network performance went above 100 megabits, Rx throughput fell off<br>
erratically (on Gig-E, they observed an average throughput of 15 Mb/s).<br>
So yes ... rx really is that much worse than TCP. &nbsp;I suspect that
this<br>
is a long-standing bug that has only manifested itself recently with the<br>
advent of faster disks and network.<br>
<br>
--Ken<br>
_______________________________________________<br>
OpenAFS-devel mailing list<br>
OpenAFS-devel@openafs.org<br>
https://lists.openafs.org/mailman/listinfo/openafs-devel<br>
</font></tt>
<br>
--=_alternative 006D449BC125721B_=--