CPU / Designer ratio

Newsletter · Original Contribution Joined: Dec 08, 2003 Posts: 1107

(Originally from Issue 2.5, Item 10.0)

From: Anonymous

We are a group of 15 asic designers with a sparc workstation for each.
Our sys admin wants to upgrade to a sparc server to decrease
maintainance effort and increase performance. I need any info about
experience with sparc servers for simulation or synthesis. How many
processors and how many memory will we need for 15 designers?

Newsletter · Original Contribution Joined: Dec 08, 2003 Posts: 1107

(Originally from Issue 2.11, Item 8.0)

From: David Morton

There can be both benefits and drawbacks to such a move. Your sys
admin is right, it will make his life much easier at least in the
short term and there can be big benefits for you (designers) and the
company. However, don't do it without careful consideration, all
large systems and client server networks address these same concerns.
As an example, some years ago a large investment bank spent
$40,000,000 to duplicate their mainframe system at a separate location
in case of failure. Their business case showed it would take only 4
hours of downtime on their main system to recoup the expense such was
their business volume on the system. So work out the costs for your
company, building on my examples that follow.

The first trap is single point of failure. If you have a single
server then when it dies (this is NEVER if, with system!), nobody
works and this can be a high cost. For example, if you have 15
designers on $125k p.a. average for 2000 hours p.a., that's about $950
per hour of downtime in salary. However, if you are selling those
skills at $250 per hour that's $3,750 income, lost forever, for every
hour of downtime. Don't forget on costs, and if you're working 60
hours weeks to meet a deadline then it's 50% more expensive. If the
whole company, (managers, accounts and secretaries) also use the same
system for their applications the cost is greater. The real expense
though may be company prestige if your email and web site are down
also.

At present, if a workstation dies, at most only one person stops
working. In reality they can probably keep working by desk hopping
when a colleague is away from their desk, so the cost to your company
may be minimal, but there is still a cost in time lost moving around.
I'll let you rework the above example in this scenario, as you know
your company best. In view of this, maybe a pair of servers would be
better, especially clustered to act as one.

The next thing to consider, is what you raised, scaling of the system.
I suggest that you find out the specs of your machines, i.e. count
CPUs, RAM & disk across the network. I would take the sum of the
number of CPUs times their speed (e.g. 15 * 366 + 12 * 750 = 14490)
and aim for half that sum in the servers. RAM and disk could be
scaled back by 25-50% because the reduced duplication and you would
still be unlikely to crash into partition and other limits because of
the greater capacities in the server. Once you have all these sums,
get boxes with room to grow. For example, your sums might show that
you can just squeeze everyone onto 2 Sun Fire 3800 4 way machines. I
would suggest looking at a pair of Sun Fire 4810s to allow more
growth. I am looking at my own EDA firm and expect to use a Sun Fire
4810 with 4 CPUs with a second installed within 2 years, growth
permitting for the very reasons I have mentioned. After the second
system comes on board, I will install additional CPUs as needed.

While this may sound negative, it is intended to get you thinking
about how to put in a better network, not just move from workstations
to servers. I want you to implement the most rigorous examination you
can, so you clearly understand what you are getting; even great
engineers can make a poor decision due to a lack of knowledge.

If you move to a server centric network, I would also look at putting
PCs on every desk with X.11 simply because for all the company's
faults, it is very hard to beat MS-Office & Windows for documentation,
email, internet and so many other things. You may even find it
worthwhile to run simulation and synthesis on the PCs. My own EDA
firm, if it goes ahead will do exactly this using laptop PCs with
external screens and keyboards, but then my staff will probably be
very mobile with a need to remotely access the EDA system. If yours
aren't mobile, desktops will save money.

Good luck and contact me direct if you want questions answered. I
spent 15 years doing this kind of stuff for people and much as I did
not really enjoy it, I became an expert.

Newsletter · Original Contribution Joined: Dec 08, 2003 Posts: 1107

(Originally from Issue 2.12, Item 2.0)

From: Simon Bates

And of course, if its not stating the obvious, you will need some load
sharing software, to allow fair access to the server(s). We use LSF
from Platform. This is a good product, as far as it goes, but what the
market needs is a good integration of LSF and Flex. This is because
you want to balance the access you give a user to the server(s) with
his access to licenses.

A non-technical issue to appreciate is that when an engineer has a
machine on his desk, he knows that what he can and can't do is
basically down to him and how he organises his work. When you
consolidate all the resources for n engineers into one place with say
n times as many Mips as was on each desktop, then the user will see n
times as much capacity than before. What he can now get done can vary
from nothing to n times what he could before and it is no longer just
down to him, it depends on what other users are doing and what access
management strategies are in place. For project managers therefore the
predictability of individual users can get worse, even if overall the
organisation does achieve better throughput.

Newsletter · Original Contribution Joined: Dec 08, 2003 Posts: 1107

(Originally from Issue 2.12, Item 2.1)

From: Frederick Hinchliffe 2nd

I can't comment on the number requirement, but I can comment on how to
minimize that number (or maximize the work accomplished) and achieve a
significant advantage over the one engineer/one computer paradigm. I
represent a software vendor (Platform Computing - product is LSF); I
will not make a pitch - you can get the facts from the web site
(http://www.platform.com).

Distributed Resource Management software intelligently assigns design
jobs to the appropriate available CPUs based on many factors (priority
of job, engineer, project; application; time of day; availability of
application license, etc.). This usually greatly increases engineer
productivity and equipment/software utilization in an organization.

The concept is well-proven and should be a part of any server-based
strategy. Look for multi-platform and OS support, scalability,
reliability, experience, support for parallel execution,
job-preemption, failure recovery, and other productivity features that
get the important work done first, but get it all done as soon as
possible.