( ESNUG 308 Item 7 ) ---------------------------------------------- [1/20/99]
Subject: (ESNUG 304 #6 305 #11) Pros/Cons Running EDA On SUN Compute Farms
> We made the change, and I like it a lot. Before, when we needed to run
> a big job, it was a mess, both network and CPU-wise. Now we have one big
> iron: it's as if you had the whole system for yourself. Of course it has
> to be dimensioned very carefully otherwise everybody will be slowed down.
>
> The problem: any downtime and your whole department is on holidays. But
> it's so much easier to administer too. But I'm not saying it's cheaper
> (it probably isn't). One of the drawbacks is upgrade: a powerful
> configuration today will seem low-end in one year. It's easier to replace
> workstation after workstation, but changing the big iron is not always so
> cost-effective. Try to negociate a good upgrade deal.
>
> - Alain Raynaud
> Mentor Graphics Meta Systems Division
From: [ The Cisco Kid ]
John - Call me "The Cisco Kid"
In my group here at Cisco we use a system based on LSF from Platform
Computing. Most of the engineers have a single PC CPU on their desks with
two monitors, running Reflection X server s/w. To do the compute-intensive
tasks, we have a compute cluster set up with about 35 total CPUs (in 18
machines), as well as after-hours access to the remaining high-speed
desktop CPUs. The users submit jobs to a queue, which has very
fine-grained control (hours of operation, number of jobs per user,
grouping, etc.) and it is then started on the appropriate machine.
Users can specify many different parameters to control which machine gets
the job (memory requirements, run time, swap needed, tmp space needed) and
all output can be either directed to a file to the user's email. It does a
great job of smoothing out high-demand periods (sometimes the jobs may take
overnight to run) without requiring any input from the users. If there are
high-priority jobs that need to be run, the administrators can either bump
them up in the normal queue, or a high-priority queue can be set up to push
those jobs through faster. Also, there is an "interactive" queue which
users start interactive jobs on so they don't have to find a machine and
log directly in to it.
Previously, we had a few compute servers that were used interactively,
where users had to hunt around for one that wasn't being used then start
their jobs and hope that no one else came in and started another one,
killing the performance. Also, each engineer had two machines on his desk,
a PC for administrative work (Word, Excel, PowerPoint, etc.), and a Unix
box to do their "real" work.
The users that have switched love it. The extra desktop space allowed by
using two monitors really helps when you're editing lots of text files, and
Reflection has a virtual desktop manager so you can have multiple
"desktops" - kind of like CDE or olvwm. The managers love it, since we
don't buy one CPU for the compute cluster for each engineer, and the ones
we buy are used more efficiently. They also like it since they can easily
monitor usage and base purchase decisions on current usage and future
requirements - it's a lot harder to get that kind of information from a
non-managed cluster.
Overall, I don't think I can say enough good things about LSF - but don't
tell them that, they'll just want to raise their prices.
- [ The Cisco Kid ]
|
|
