Comparison of the efficiency of SMP cracking vs. node-based cracking (talking RC5 here)

[beer]

OK - I've been thinking about this and I'm going to *try* and not sound like an idiot here. If I do, trout-slap me. So here goes...

Let's assume for the duration of this message that you are running a Linux-kernel based machine. Basically, which ones of these would be most efficient?

1) An Abit BP6 with dual Celerons, clocked 366 -> 550, running in SMP running RC5.

-or-

2) A small cluster - assuming you had an adequete network server and topology, with two nodes, running single Celeron 366s at 550.

Now, assuming you were to run, say, an 8-way Xeon server in SMP. Since RC5 is entirely based on CPU horsepower and not cache, would you see better performance in either -

1) An 8-way Xeon server running in SMP, at, say, 550 MHz a piece.
2) A cluster with 8 nodes, Celeron 550s, and adequete server and switching/network topology.

I'm trying to think of the scalability here....I'm not too familiar with node-based multiprocessing now...

 

[jatwell]

If you're talking pure RC5, smp boards are better.

That way, one processor doesn't have to deal with the overhead of the OS. It's almost not worth worrying about.

I would imagine the Xeon would eek out a win by a couple of blocks.

 

[Kilowatt]

Good question Elemental007.

You'd have to look at this way, an 8 node single processor setup will have to spend a little time talking to it's server through the network, taking a small fraction of it's cracking time, say a couple seconds every so often.

SMP, on the otherhand would be a few "nano secounds" if the board had it's own HDD with plenty of blocks.

The thing about SMP, is that the OS overhead, is split between the CPUs. Load balancing so to speak.

The most efficient, would be to use Dual boards for your Nodes, with dual Ethernet Channel Bonding, but you won't see that much of an improvement in RC5 to justify the cost of extra components.
You would see an improvment in calculating the circumferance of a galaxy, or ploting the track to get you there, something that might take 30 days on a large Beowolf.

Booting diskless, with my Klinux, for example, reduces the overhead you will encounter with a full blown OS installed.
Granted, that Linux is much better at saving it's system resources, than Win32, you can (i have) forget to turn certain things off in Linux, and waste cycles that are better spent cracking.

Even an 8 CPU Beowulf will not have the faster keyrate (all be it a small differance in the overall picture) over a system with all 8 processors on the same motherboard, all due to a little network traffic.

A CPU cycle is a terrible thing to waste, but so is the big $$$ on a 8 way motherboard. (unless you get one real cheap )

I'm in the process of building my own little "miniwulf", and decided to use 4 BP6s since I already had them.
The RC5 numbers are about right for 4:1 ratio over a single BP6 running the same CPU's.

OH Shoot, did I get sidetracked?
What was the question?

Oh Yeah, two single mobos running a 550 cel, will get the same keyrate as a single BP6 running dual 550, providing their running the same OS.

 

[Dale]

<<Basically, which ones of these would be most efficient?>>

efficient in terms of cost per key ??
efficent in terms of space ??

the output is basically the same

it is a lot easier, neater to setup 1 BP6 than to do 2 ??
less boxes,cards,cables,hubs,etc

it is mostly a $ decision

I only have 1 BP6, wish I did have 4 ; )
mine has been running since the first week they were available
and is only turned off to rob parts from it : (

..Dale