A very, very crazy mac mini cluster idea...

I've been in the market for a new computer for awhile now, preferably a mac. I was thinking about getting a Mac Pro, however then I had a wacky idea today (and it's past 4am while I write this, so maybe this is just wacky period...). I was comparing costs and I could put together 5 mac minis for the cost of a single Mac Pro upgraded to 2.5 gb of ram. I did a little googling on the matter and found out some people had done similar things. Here are the spec comparisons:

5 Mac Mini:
5X 1.66Ghz Intel Core Duo --> theoretical max 16.3ghz??
5X 512MB DDR2 667 memory --> 2.5gb just like the Mac Pro
5X 60GB SATA hard drive --> 300GB

Mac Pro:

Two 2.66GHz Dual-Core Intel Xeon "Woodcrest" processors --> theoretical max 10.64ghz??
2.5GB memory (667MHz DDR2 fully-buffered DIMM ECC)
NVIDIA GeForce 7300 GT graphics with 256MB memory
250GB Serial ATA 3Gb/s 7200-rpm hard drive1
Looking at the raw numbers, it would appear the mac mini cluster is a better deal. Of course I oversimplified above and I know a Xeon is not the same as a Core Duo, however I'm interested to see what others think about the potential here. If I go the Mac Mini route, I'll be waiting around 1 year until they make it quad core and beef up that anorexic hard drive. I might consider going 1 gig on all of them for 5 gigs of ram.

I wouldn't be using it for anything "seriously server optimized" so perhaps the Xenon/Core Duo difference is negligeable?

Forget about it. Unless you plan on doing Computational Fluid Dynamics as a hobby, any cluster setup would be worthless for you (and if you were doing CFD there are better choices than a bunch of clustered Mac-Minis.)

Here's the deal:

If you want a general purpose system that plays SL very well, get a high-end iMac (20" or 24" inch.)

If you're the kind that must have the ultimate SL graphics performance, top draw distances, all features turned on, etc., then get a PC with a monster GPU (first choice), or in the Mac world, get a Mac Pro (second choice, by some distance behind.)

As for Xeon vs Core Duo, for most purposes a Xeon system is *inferior* to a Core Duo system. Unfortunately most people who buy the Mac Pro do not understand this (and unfortunately the Mac Pro is the only Mac platform you can use a top-end GPU.)

Xeons are basically a bad choice for anything. The Opterons, on the other hand...

-peekay

Thank you very much for the reply Peekay I'm glad to know the Xeon isn't better than a Core 2 Duo. I might wait and see what the next Imac upgrade brings... The Imac in black that I've seen rumored on the net is mighty spiffy

Many Xeons can be better than one Core Duo, if the application can take advantage of it (most real-time 3D apps can't.) It could be an excellent choice if you do heavy multimedia work, for example. Then again, them Opterons will do even better in those situations, at less cost.

I think a 24" black iMac would be totally sweet.

-peekay

you are totally insane... how can you even say that the Xeons in the Mac Pro are inferior? Are you assuming they are the same Xeons that were P4 based? they arent.

The Xeon 5100 series used in the Mac Pro are the same exact core design as the Core 2 Duos. They are basically Core 2 Duo processors that run a faster bus, faster core speeds, and can run with multiple sockets. The only thing inferior is the chipset that requires FBDIMMs, which arent really as good as standard DDR2 memory for normal day to day desktop use, but for that type of operation is not really going to be noticeable... except in the pocket book.

We've benched our own (data processing) application on both processors and found the per-core-mhz Xeon performance lags behind the Core 2 Duo, precisely because of the inferior memory configuration.

You can't process data unless you can get it to the CPU and while FB-DIMM specs look ok on paper, in reality with current chipsets the Core Duo's memory subsystem is roughly twice(!) faster than the Xeon's. That is a significant difference.

As an example, our app's first phase is processing about 14gb of compressed satellite data, and on a per-core basis the Core Duo completes this phase 50% faster than the Xeons! Other phases are less dramatic but there are zero phases where the Core Duo noticeably lags behind the Xeon, but there are a few phases where the Xeon lags behind the Core Duo.

So I stand by my assertion that for a general purpose app mix, a Core Duo will either outperform Xeon system -- or at least tie it -- given the same number of cores. And for a much cheaper price!

Of course Xeon's advantage is being able to have more than four cores per system (except on Macs because the Mac Pro is limited to 2 CPUs.)

-peekay

Edit to clarify our tests were against the Core 2 Duo.

what Core 2 Duo, and what Xeons? what system specs, size and amount of dimms and how they were installed... Sounds like you had things configured badly. Core 2 Duos around the same clock speed, on single memory intensive apps, can slightly outperform the xeon systems with the 5000X chipset, but thats about it. If you were seeing that huge of a difference, you had something wrong.

but thats beside the point... there is no other Mac sold that can outperform the Mac Pro. And the Xeon is not a slow processor, everything your saying is slow is a chipset and memory problem.

It's pointless to bench two CPUs without regard to its real world performance given available motherboard, chipset, memory, etc. People who purchase Mac Pros buy entire systems, not just two CPUs glued together in a vacuum.

Here's a third-party benchmark using various Woodcrest Xeons and Core 2 Duos:

http://www.gamepc.com/labs/view_content.asp?id=5160vs6800&page=1

Their results align closely with my findings above. Note in particular its memory bandwidth test, the memory latency test, and its two compression tests (which would be very similar to our app's data parsing phase.)

The Core 2 Duo based system is not only significantly faster than the Xeon system on many test (while never noticeably behind on any), but when you figure in total system price/performance, there's just no contest.

Apple would serve the community well by offering a modular Quad-core Core 2 Duo desktop as a choice. It would perform just as well as Dual Xeons (maybe even faster) at a significantly lower total system price. People look at the Mac Pro and just blindly assume "oh, it's Xeon, must be better" when the reality today says otherwise.

-peekay

It would be great if Apple did something like that, but the truth is Apple doesn't use any desktop Core 2 Duos, they only use the mobiles. even if they used quad core Core 2s they could still have 8 cores in the Mac Pro since you cant do multiple sockets with the C2s. It all about trade offs.

"Xeons are basically a bad choice for anything. The Opterons, on the other hand..."

I'd much rather pay license costs for a 2 socket, 8 core Xeon system, than twice as much for a 4 socket 8 core Opteron system, when hey are just about identical in performance. When AMD gets their new quad cores out, that could, and most likely will, change.

I agree though that Apple should make the Mac Pro be a C2D and C2Q machine instead of Xeons, but its still better performance than anything else Apple makes.

Regarding the performance issues with the FB-DIMMs, during the various tests mentioned in this thread, was the entire 256-bit wide memory bus utilized?

I know that Apple specifically recommends using at least 4 modules to properly take advantage of the memory bus in the Mac Pros... wondering if this could make for performance improvements?

My setup is 2gb in 4x 512mb sticks, which I purposely did due to the recommendation and based on what I could afford at the initial purchase. For the other four slots I will be going with 2gb chips in pairs as I can afford them.

In most cases you wont see any measurable difference between using 2 channels vs. 4 channels. And even with 4 channels, FB-DIMM today is still significantly slower than "plain-old" DDR2 RAM rated at the same MHz.

Adding insult to injury is the fact that you can get 1067 MHz DDR2 parts cheaper than 677 MHz FB-DIMM chips. We're talking twice the memory bandwidth with half the latency, for less money.

FB-DIMM might be the memory architecture of the future, and may even make sense for "server" configurations today, but unfortunately right now it's not living up to its theoretical potential.

-peekay