Networker

[Networker]

2005-06-25 08:47:41
Subject: [Networker]
From: Dave Gold - News <dave2 AT CAMBRIDGECOMPUTER DOT COM>
To: NETWORKER AT LISTSERV.TEMPLE DOT EDU
Date: Sat, 25 Jun 2005 08:47:38 -0400
I've been told that the co-processed NIC's on Solaris only reduce utilitzation 
by about 30%; it seems that the coprocess side just checks for whether a packet 
is targeted at the host or not. Something inherent in the design of the TCP 
stack on Solaris, from the explanation.
 
(Feel free to disagree...I'm just reporting what I was told...)

Dave
 
Date:    Fri, 24 Jun 2005 12:26:48 +0200
From:    Oscar Olsson <spam1 AT QBRANCH DOT SE>
Subject: Re: Parallelism???

On Thu, 23 Jun 2005, Eric Wagar wrote:

EW> Depending on the version of NetWorker will depend on how many parallel save
EW> streams you can run.  For us, we have the Network edition, so we get 64.
EW> With a library with 10 drives, we use all 64 streams on a save.  We, like
EW> others, can only use one stream on a tape based recover.

What server do you have that can manage that? What aggregated throughput
speed to you get?

We have a Sun V440, with 4 1281MHz Ultrasparc IIIi CPUs with 1MB cache,
and that server can handle just about 1gbit of throughput, before the
kernel just can't get more CPU time. We're using the built-in ce (Cassini)
NICs. It seems like the excessive CPU useage is caused by network
processing. We can't use jumbo frames, since some network equimpent, that
the clients are attached to doesn't support it.

To me, the excessive amount of CPU used to produce 1gbit of network
throughput seems just plain wrong, since I seem to remember that the ce
chipset has TCP checksum acceleration? Or are there any faster NICs in
this regard, how about the bge chipset? Or would a better CPU (same
speed, but larger cache, and yes I know I need to get a new server
then :) ) with a larger cache do the trick, since we see a lot of context
switching going on?

The funny thing is that if I stream data to the drives alone, without
reading the data from the network, it takes very little kernel CPU.

//Oscar

--
Note: To sign off this list, send a "signoff networker" command via email
to listserv AT listserv.temple DOT edu or visit the list's Web site at
http://listserv.temple.edu/archives/networker.html where you can
also view and post messages to the list. Questions regarding this list
should be sent to stan AT temple DOT edu
=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=

<Prev in Thread] Current Thread [Next in Thread>
  • [Networker], Dave Gold - News <=