Disk subsystem effects on performance?

[May Kung]

This subject came up at work of late, whether for a given amount of money, it makes more sense to buy a faster CPU, more RAM, faster HDD (SCSI vs. IDE?), faster graphics card, etc.

Now, it's my understanding that for general 3D modeling purposes, the faster CPU and RAM probably play a larger role than the graphics card, though when manipulating large models, the card certainly plays a role.

Since OSD uses the HDD to temporarily store and expand a model when loading/saving, it would make sense to use a faster SCSI drive vs. an IDE drive. I can see the difference when running SD9, but I'm wondering if anyone else has noticed this on OSD?

[clausb]

I'd question your assumption that SCSI disks are faster. Today, you get a MUCH better bang/buck ratio when buying IDE disks. Only in certain circumstances can SCSI disks significantly outperform IDE disks, for example in systems with several disks where multiple transfers are going on all the time, i.e. typical server configurations.

In real life, I have often enough seen IDE drives outperform SCSI drives. I would always opt for the IDE drive if given a choice. You'll save enough money to be able to a) get a much larger drive and b) upgrade earlier in the future when you find that you need a still larger and faster drive and c) buy more RAM (because this will reduce paging activity to disk, which is one of the biggest performance problems, no matter which disk technology).

There's one caveat: Always make sure that the IDE drive is configured to run in DMA transfer mode (rather than some CPU-bound PIO mode). If you're unsure whether the IDE disk is configured correctly, try tools such as DMACHECK (see, for example, http://www.spacewalker.com/english/support/faq/top10QA/WinNTDMA.htm for a description).

The choice of graphics card does not only influence interactive viewing performance. If the graphics card supports occlusion culling in hardware, you will also often see a 2-3x performance improvement in Annotation view updates for assemblies. Certain operations in the Simplification module will also benefit from occlusion culling support in hardware.

Claus

[May Kung]

I don't deny that IDE drives offer much better bang for your buck. For the average user, SCSI drives are overkill. Typically, we order workstations that already have a very fast CPU (currently, using 2.8 GHz or so), a Wildcat card, and 1 GB of RDRAM. Considering all of the above, and since the motherboards Dell uses on this line have onboard SCSI controllers, it doesn't cost much more to use SCSI disks, especially if capacity is not an issue (everything is supposed to be put on the network).

If the choice is more RAM or a SCSI drive, then by all means, go for the RAM. It makes a lot more sense. However, if you already have 1 GB of RAM, and a fast card, and a fast CPU, what is there left? Does it make sense to spend that money towards another GB of RAM, or opt for a SCSI drive instead?

I've seen SD9 running on IDE vs. SCSI and do notice a difference, but it could be just my perception, which is why I'd like to see if there is a possible explanation for this.

[John Scheffel]

Whether or not you will benefit from more than 1 GB of RAM really depends on the size of the models you are working with. Some people who work on large assemblies will start swapping to disk even with 1 GB. It's not too hard to tell when this happens, especially if you have a noisy disk. You will notice a lot of disk activity while working on you're model and it will be pretty sluggish. If you don't have enough RAM to avoid swapping to disk with the size of models you typically work on, then you would be better off putting money into more RAM than any other factors.

[May Kung]

Is there a factor for estimating the amount of RAM based on the model size? We've used 10:1 in the past (e.g. 50 MB file would need 500 MB free). For the users that deal with larger assemblies, going from 512 MB of RAM to 1 GB made a big difference. We haven't tried bumping it higher, though. I'm not sure if some of our workstations can handle more than 1 GB of RAM. Some definitely can, but that'd require pulling out RIMMs because all the banks are full.