Commit 70b4e741 authored by Mike Hibler's avatar Mike Hibler
Browse files

Quicky description of good, bad and ugly of using pc3000s.

parent c2f2c858
Copyright (c) 2005 University of Utah and the Flux Group.
All rights reserved.
<h1>The New "pc3000" Nodes</h1>
<li> <a href="#machines">Machines and Interfaces</a>
<li> <a href="#images">Images and Kernel Support</a>
<li> <a href="#caveats">Caveats</a>
<a NAME="machines"></a><h2>Machines and Interfaces</h2>
The new "pc3000" machines are
<a href="">
Dell PowerEdge 2850s</a>
with a single 3GHz processor, 2GB of RAM, and 2 146GB SCSI disks.
Each has 4 available Gigabit experimental net interfaces,
however due to lack of Gb switch ports, not all are usable at Gb speeds.
Instead, they are a mix of 100Mb and 1Gb as follows:
Machines Gb ports 100Mb ports
40 4 0
20 3 1
40 2 2
60 1 3
Note that all 160 machines have at least 1 Gb interface and all can be
used for 100Mb as well.
There is also a 5th Gb interface on each machine, that is used to pairwise
connect adjacent machines. For example, pc201 is connected directly to pc202
via a cross over cable. These interfaces are not yet available, pending
further support in the resource mapper.
All machines have serial console lines accessible from via
the "console" command or remotely via the
<a href="">tiptunnel</a> interface.
<a NAME="images"></a><h2>Images and Kernel Support</h2>
Currently, the Emulab standard FreeBSD (FBSD410-STD, FBSD54-STD),
Linux (RHL90-STD, FC4-STD), and Windows (WINXP-SP1-pc3000, WINXP-SP2-pc3000)
images will run on the new machines.
If you have built custom images based on our standard images before
Sept 2 2005, they will likely not work on the new machines due to a lack
of the correct disk driver. You will either need to re-customize based
on the current images or modify your existing image to add the correct
SCSI driver. For BSD you need:
device mpt
in your kernel config file, and for Linux you need:
in your .config.
For Windows you will need to re-customize based on our current -pc3000
<a NAME="caveats"></a><h2>Caveats</h2>
<li> <b>No bandwidth delays between 100Mb and 1000Mb.</b>
We have not yet looked at providing bandwidth delay values between
100Mb and 1000Mb.
<li> <b>The "direct connect" 5th interface is not usable.</b>
As mentioned, we need modifications to the resource mapper
and there are also security issues to be resolved.
<li> <b>Interswitch bandwidth may limit topologies in non-intuitive ways.</b>
Since the Gb interfaces and the 100Mb interfaces are on different
switches, and the new machines are on different switches than the old
pc600 and pc850 machines, the bandwidth of the interswitch links can
have a significant impact on whether an experimental topology can map
or not.
<li> <b>No vnode characterization has been done.</b>
The so-call "colocation factor" has been arbitrarily set to 75 right
now. It will likely be clamped down considerably.
<li> <b>There might be assorted failures due to scaling issues.</b>
Experiments of over 150 machines, which were hard to do before,
may reveal further scaling issues that will need to be addressed.
Expect an increase in unexpected swapin failures as a result.
<li> <b>These machines are radically different than the old machines.</b>
Since these machines are so much faster, have so much more and faster
memory, have so much more and faster disk, results of your experiments
on these machines could be dramatically different. That is, if your
applications were CPU, memory, or disk intensive. If you are
interested in reproducing old results, you might want to
<a href="">
limit your experiments to using only pc600 and pc850 nodes</a>.
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment