RP3440 vs. RP7420

[Carl Houseman]

I'm trying to make an apples to apples comparison of these two hardware platforms, The goal is to acquire a "hot spare" system (RP3440) which can stand in for a number of days in the event of serious trouble with our RP7420.

So given the following:

(A) RP7420 with 5 x 900Mhz cores and 24GB allocated to a vPar running Oracle 9i database
vs.
(B) RP3440 with 4 x 1GHz cores and 24GB running Oracle 9i database

BOTH: dual A6795A 2Gbps FC adapters to a VA7110 disk array.
BOTH: system on AD188A U320 36GB 15K drive
BOTH: HP-UX 11.11

In production, the RP7420 rarely needs all 5 processors. The performance is usually constrained by I/O to the VA7110 array.

The $64M question:
Can I reasonably expect the RP3440 to perform to at least 80% of the RP7420's ability? Or even 90%?

thanks all...

[Torsten.]

Hi,

you said "5 cores", but I guess it must be 5 CPU modules = 10 cores.

The rp3440 has only a maximum of 2 modules = 4 cores.

Hm, 4 < 10, as you know.

So I guess you unlikely will get 90 %.


Next problem:

rp7420: 16 GB/s aggregate I/O slot bandwidth

vs.

rp3440: 2.5 GB/s aggregate I/O slot bandwidth (rp3440)

and another:

rp7420: 2 to 128 GB memory capacity

vs.

rp3440: 32 GB memory capacity


So I would try to compare "real" performance measures or ask the DB experts in the hp-ux forum next-door here on itrc.

[Carl Houseman]

No, when I said "cores" I meant cores, not modules. I know the difference. Glance will show 400% MAX on the RP3440 and 500% MAX on the RP7420 vPar.

And when I said 24GB memory I meant 24GB memory, period. Apples to apples. I don't care that the RP7420 can expand to 128GB - both are limited to 24GB for the purpose of this comparsion question.

Since my first post I have found that the peak memory bandwidth in the RP7420 is 16GB/s PER CELL BOARD (this RP7420 has two cell boards and other processors in another vPar). The RP3440 memory bandwidth is 8.5 GB/s. What I don't know is how much memory bandwidth might affect bottom-line performance.

And as for I/O bandwidth, I really only care about 1 PCI-X slot used for the 2Gbps FC card on the primary path to the VA7110. Even the lower-speed slots in the RP3440 can saturate a 2Gbps FC adapter. The fact that many more slots and much more I/O bandwidth are possible in the RP7420 - don't care. Apples to apples.

I will take the suggestion of posting in another forum, but I think in this forum, the question boils down to memory bandwidth. Anyone who wants to chime in with agree/disagree on that feel free!

[Torsten.]

OK, I missed the "vPars" related to the cores - this sounds better now.

[rick jones]

While the rp7420 does indeed have higher aggregate memory bandwidth (since IIRC you have two cell boards, each with two memory busses) the extent to which you see that full bandwidth depends quite heavily on how fully populated the DIMM slots happen to be. You really want to have as many of the DIMM slots filled as possible in the sx1000-based systems.

The rp3440 probably has lower latency. You could check that by running some lmbench tests.

If you are indeed usually bound by the one 2 Gig FC link to a VA7110 I'd bet that the rp3440 would remain the same.

Best thing to do of course is try to get some time in one of the capacity planning centers (now called Solutions Centers IIRC) and actually try-out your workload.

[Gavin Hubbard_1]

I don't know the specifics of your application, but I would expect a VA7110 to be a major bottleneck for either of the systems you mentioned.

[Carl Houseman]

Yes, VA7110 is a bottleneck but that is IRRELEVANT to this question.

[Gavin Hubbard_1]

Why is it irrelevant? If your application is I/O bound then there's going to be little or no difference in performance between the 7420 and 3440.

Having said that, without knowing any details about your application it isn't possible to make a meaningful comparison between two different hardware platforms.

[Carl Houseman]

The application is I/O bound much of the time but not all of the time. In this ocmparison I/O performance remains constant. I want to be reasonably certain the CPU performance is not degraded too much due to internal architecture differences of the RP3440 compared to RP7420.

[Carl Houseman]

For Rick Jones - I appreciate the comments regarding bandwidth vs. latency, and I can tell you that all DIMM slots on both RP7420 cell boards are populated.

Since you seem to have some knowledgeof RP3440 architecture, is there a performance difference for populating 3 quads (to 24 GB) vs. 2 quads (to 32GB)? I would think 3 quads should increase the aggregate bandwidth. True/false?

Solution center is not an easy option. Short of putting together a system, transferring production to it, and telling our users to do their workload for a week, every attempt at quantifying an answer through simulation will be no more reliable than attempting an analytical answer.