L1000 Upgrade

[Dave La Mar]

We have a L1000 that for monetary reasons will cost more for a second processor than to upgrade the 440 to a 550.
Has anyone experienced this upgrade and the added punch results?
Obviously two 440's sounds better than one 550 but I need some input since costs involve sotware increases based on number of cpu's.

Any and all input appreciated.

Regards,

dl

[Ted Buis · Accepted solution]

Well, the greatest performance increase you will see going from 1 440MHz to 1 550MHz is 25%. Since memory speed hasn't changed or RAM hasn't been increase then that is no gain at all in those areas. Are the disks going any faster? Where is your current bottleneck? Is it CPU or network or disk or RAM? In other words, without knowing more, your gain could be 0 to 25%. The second 440MHz chip would likely deliver more performance but that too depends. It won't if you are I/O bound.

What I would suggest is that you look at the rp3410 if you can run your application on 11i, since then you get to a 800MHz processor. These new systems don't cost that much but will get you a bigger bang for your buck if your main cost is from per processor software licenses.

[Dave La Mar]

Ted -
Good input, but I really do not want to buy a new machine.
The bottleneck is in both memory and cpu.
The plan is to add 1 gb memory, but the cost of a second 440 is out of line due to # cpu based software products.
If we gained 25%, that would be acceptable.
I am looking to hear from someone who did just this, success or failure.
I value your comments and will respond with points when ready to close this thread.
Thanks again,

dl

[Ted Buis]

If you knew the instruction and data average cache hit values, then you could estimate peformance fairly well. At zero cache hits, to would get zero improvement. My thought is that if you had a 90% cache hit average you would get 90% of that 25% clock rate gain. Unfortunately, CXperf was the tool that could get you that number, but I think it is no longer supported, and Calipher is only supported on Itanium systems to the best of my knowledge. I didn't think you could get that from Glance, but I'm not an authority on this. If you do install the higher clock rate chip, and there was no I/O, then you could deduce what you average cache hit value was. ;~)

[Dave La Mar]

Well, since glance was not licensed on that machine, we just took a top snapshot on a .5 hourly basis for a period of time.
This clearly indicated we were pushing th one cpu.
Since there is no additional cache benefit going to the 550, we'll probably just bite the bullet and get the second 440.
Thank you Ted for your input.

Please consider this thread closed.

dl

[Ted Buis]

Since the L1 cache on the 550MHz pa-8600 is on the chip it runs at the processor speed, so you do get full performance for all instructions and data that are in the cache, but when there is a cache miss, then the fetch of the cache line takes exactly the same amount of time regardless of the clock rate, since the bus speed from the processor to memory and the memory speed itself doesn't change. Hope that was clear before this, but I just wanted to make sure I hadn't given you the wrong impression. Since all the data and instructions won't always be in the cache, the gain will always be something less than the 25% clock rate gain.