Re: [Nbd] Performance tuning of nbd !?
- To: Dirk von Suchodoletz <dsuchod@...65...>
- Cc: nbd-general@lists.sourceforge.net
- Subject: Re: [Nbd] Performance tuning of nbd !?
- From: Paul Clements <paul.clements@...124...>
- Date: Fri, 29 Feb 2008 12:08:33 -0500
- Message-id: <47C83C11.8060401@...124...>
- In-reply-to: <Pine.LNX.4.64.0802282217000.5163@...163...>
- References: <Pine.LNX.4.64.0802282217000.5163@...163...>
Dirk von Suchodoletz wrote:
When I try it over cable (okay, might be a little bit perverted - 25Mbit/s
ping of 15ms) or DSL (14Mbit/s ping of 35ms) the results drop to less than
200kByte/sek. It is still possible to run a stateless client that way
(squashfs ontop of nbd), but you have to be patient while an app is
starting the first time.
Is there any chance to get above the 10Mbit/s LAN rate in the WAN DSL or
cable scenarios!?
Yes. I think you should be able to use about 70% of the bandwidth (I can
get about 32Mb/s out of 45Mb/s (T3) with 100ms latency). You might need
to increase your TCP window size and/or change the I/O scheduler you're
using for NBD. Switch to "deadline" scheduler if you're not using that
one already. And set the TCP window size, as follows:
In order to raise the window size to about 1 MB, the following sysctl
settings should be changed:
net.core.rmem_max = 500000
net.core.wmem_max = 500000
net.ipv4.tcp_rmem = 4096 87380 1333000
net.ipv4.tcp_wmem = 4096 16384 1333000
The window size should be a little above your network "pipe size":
pipe = bandwidth * latency
--
Paul
Reply to: