Commit 7ee620b3 authored by Mike Hibler's avatar Mike Hibler
Browse files

New project: power saving, updates on tip and disk loading

parent 58ac59e8
......@@ -8,6 +8,8 @@ Things that could be done:
2. Filesystem compression (mike)
[ Leigh is working on this. ]
Jay's Holy Grail of Ultimate Disk Image Compression. I see three
ways of doing this:
......@@ -85,6 +87,9 @@ Things that could be done:
3. A new tip (mike)
[ Mike fixed the easy stuff: ripped out most of the escape sequences
and dialer support, reduced to one process, cleanup tty on exit. ]
Things that are wrong with tip in the environment we use it for
(remote console access):
......@@ -143,9 +148,62 @@ Things that could be done:
6. Reduce (or eliminate) logging in default OSes
The standard daily/weekly/monthly activities are occuring and
The standard daily/weekly/monthly activities are occurring and
sending mail to root (at least in our FreeBSD image). We should
either eliminate this or, better, reduce it and have it logged
on plastic or paper. We should at least do some security checking
and mail any anomolies to testbed-ops or someplace so we know if
and mail any anomalies to testbed-ops or someplace so we know if
some machine has been cracked.
7. Reduce power usage
We should do what we can to reduce excess power usage. The most
obvious thing to do is turn machines off when they are not assigned.
The benefits are:
- maximum power savings
- reduced vulnerability to script kiddies
- longer component life?
- no rc5des!
The downsides are:
- can't use unassigned machines for other purposes (abone)
- don't save anything if the testbed is in use 24/7/365
- frequent transitions would be harder on the HW
At the very least, we should make our scripts capable of dealing
with this model. For example, os_setup would need to check and see
if the machine is powered down and, if so, turn it on. Likewise,
deassignment should turn the machine off. Note we will need to
differentiate "down to save power" from "down cuz it emits large
billows of black smoke when turned on." Autostatus would have to
be taught to show "available/assigned/down" instead of "up/down".
To avoid excess transitions due to frequent reassignment, we can
delay shutting down a machine "for a while" after an experiment
has ended. The end-experiment algorithm might look like:
if someone is waiting for an available node,
immediately reassign/reload this one
else
"clean" (reload/zero) the disk
if the node still isn't needed,
power it down
That would give a 5-10 minute window for reclamation.
What can we do for machines that are turned on? For assigned
machines we probably can't dictate much. The best we can do
is make our default kernels "green". Both Linux and FreeBSD
probably idle the CPU, so we are probably ok there. Not sure
about spinning down the disk. Because of daemons and logging it
might be difficult to keep the disk spun down. Dave/Kevin/Steve
worked on this some on the Jaz-disk test machines. Some keys are:
- shutdown as many daemons as possible
- reduce logging and log across NFS if possible
- move unimportant logging to a memory disk
Markdown is supported
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