BUS Performance differences between equally configured N & L class

[Edward M. Hoppe, Jr.]

I'm looking for an end-user perspective (hopefully through experience) system performance differences between a dual-L2000
with 440MHz CPUs and an equally configured N4000 - dual 440MHz
CPUs. Both servers have 6GBs of RAM and 2 (two) 440 MHz CPUs. I was curious if anyone
has migrated from a L2000 to the N4000 device (all else being the same) and if so, what (if any) were the performance enhancements. I'm actually going the other way (from the N to
the L) in a development environment so I can free up a device (the N since it has more expansion capabilities)for my
production environment. Any performance comparsion info would be most appreciated. I don't want to leave my development folks with a 'dog'! since it's an Oracle database server with quite a few instances defined. The back-end disk is also the same -using SAN storage served up via the XIOtech Magnitude - same RAID level and number of drives / speed allocated for the databases. Tried to keep things as equal as possible, yet wait IO seems to be quite higher on the L vs the N.

[James Murtagh]

Hi Edward,

Unfortunately I can't give you too much from an end-user perspective, don't do to much real work anymore! However, if you haven't seen the following white papers on the two server classes they should prove very useful. They describe and illustrate the system architecture and detail the respective bandwidth of various buses. The main difference you will see if that the N-class has two system buses and twice the IO throughput of the L-class. More information can be found on the main hp website and selecting the servers link.

http://www.hp.com/products1/servers/rackoptimized/rp7400/infolibrary/rp7400techwp5.pdf

http://www.hp.com/products1/servers/rackoptimized/rp5400series/infolibrary/pdfs/5400-whitepaper.pdf

Cheers,

James.

[Brian M Rawlings]

Hi, don't mean to dredge up an old post, but I thought I'd add something to this in case it benefits some future data miner down the road.

No points, please.

The answer is (as always), "it depends".

If you are in a heavy I/O environment, with enough FC or Gig-E cards to actually stress the I/O buses (hard to do with only two CPUs, in either box), you would find the N-class (RP7400) to have a clear edge.

As noted, the N-class has two major bus structures, which allow for ideal "pairing" of disk I/Os when using two host adapters for each array connection. What the N-class units have that the L-class (RP54xx) does not, is that all but two of the PCI slots are "Twin Turbo", 4x the original PCI throughput. Each slot has its own PCI bus/chips, unlike most other servers (including the L-class), where PCI slots share the PCI bus/chips.

This gives the N-class monster I/O capability. I suspect the N-class design team overcompensated for the doggy performance of PCI in the V-class, where a 7-slot PCI chassis had two PCI buses/chip sets, 3-slots and 4-slots sharing one bus. So, it is basically impossible to bottleneck I/O on the N4000, at least where PCI speed is concerned...

But that is only part of the story, the other part is memory latency. Long story there, but the short version is that the more CPUs you have, the more layers you have to add to your memory bus & access design in order to let all CPUs share all of memory.

Since the L-class only goes up to 4 CPUs, it has lower memory latency than the N-class, which can go up to 8 CPUs. For identical speed CPUs (in identical numbers), this translates to better overall speed for the smaller box.

The same is true as you move further down the scale to the A-class (RP24xx), which only has to design around two processors. In the same CPU speed, and with all servers dual-processor, the A-class is the fastest, the L-class is next, and the N-class is last.

These results bear out if you look at the TPC ratings for each server (www.tpc.org). This is a little surprising, since the TPC-C is an I/O intensive benchmark, but all it means to me is that the testing never seriously stressed the I/O side, to the point of bottlenecking on disk I/O or Client I/O (the TPC-C is a client-server, multi-tier benchmark, requiring large databases and lots of network traffic). If it had, I doubt the N-class would have been last. As I said earlier, two CPUs can't pump enough I/Os to make its I/O structure sweat at all.

Once you add enough CPUs to make it interesting, the A-class cannot do an apples to apples comparison, since it stops at a 2-way. Even the L-class can't really, since the N-class probably can't pump enough I/Os to really fill its big I/O pipes until you get 6 or 8 CPUs running, and the L-class stops at 4.

What you will find is that your developers will be very happy with the performance of the 2-way L-class, if production runs adequately on a 2-way N-class.

Then, of course, there's memory interleaving to consider...

Don't get me started.

Regards, --bmr