Gigabit Ethernet over Cat 5 cable

[Guest ebedding]

Has anyone looked at or tested the new 10/100/1000 network cards now available that work with standard Cat5 cable. We all work off a central server and I am looking into reducing the IronCad save and load times of large assemblies.

 

The cost of 10/100/1000 network cards are now around $110 ea and Gigabit Switches start at $600 for a 3COM 6 port and go up from there.

 

Would a decent server with 15,000rpm hi-speed SCSI drives and a Gigabit network card result in significant faster save and load times in IronCad for the end user at a standard workstation also using a Gigabit network card vs 10/100?

 

 

[Chris Lohman]

Eric, here's my 2 cents. I write this with achieving the highest possible read times for loading IronCAD files only, in mind.

 

1) No matter how you load it, the client's disk is ALWAYS a factor in the speed the file is loaded. IronCAD Scene's are actually compressed files that we have to uncompress before we can work with it. No matter where you load the scene from, IronCAD uncompresses that scene file using the local client disk, and uses temporary versions of that file locally. IronCAD uses the $TEMP environment variable to determine where to uncompress those files to (as well as RAM). With this in mind, upgrading the client disk should always be considered.

 

1a) ATA verses SCSI. ATA guys lately are touting their "7200 rpm" disks and trying to tell us that SCSI is old and inferior by referring to it as "Legacy". SCSI is Legacy in that it has been around for a very long time, but that doesn't make it bad. SCSI has been constantly growing and maturing and has far more years of experience behind it than eide, ide, and ata. No matter what they tell you, as of right now SCSI is still king with disk access rates BY FAR (if you have a top of the line scsi disk of course).

 

1b) SCSI verses RAID: I was recently able to experience a system where the disks were running on the built in Ultra3 SCSI controller, and then moved into a raid0 configuration through an add-on card. These were 15000 rpm SCSI disks from fujitsu. There was a very noticeable difference in read and write times between the two; the raid was definitely faster. Even a single disk on a RAID card was much faster than straight SCSI. The RAID card had a 64meg Rambus RIMM in it and I had set up the controller to perform read caching but direct writing. I suspect the significant speed improvement was due to the controller caching the file into it's 64 megs of RAM, thus diminishing the bottleneck of disk access rates. RAM is just faster than a disk so the fact that I was able to perform read-ahead caching to the RAM on the RAID card (and thus the processor accessed the data from RAM instead of directly from disk thereby increasing my access rates to almost the speed of the system bus) is (I believe) what made the big difference. So a single disk on a RAID card with ram in it (that I could configure for caching) is much faster than a single disk on a scsi controller. In a situation where you split up your disks to accommodate what I mention in item 1c below, you would want one channel per disk and ideally one controller per disk to achieve the absolutely highest possible disk access rates. (I'm talking about the very fastest you could possibly go though, this would be extremely expensive and isn't realistic for "most" clients. If you're working with scene files that are 800 megs+ in size though it might not be that crazy. In a server environment you definitely should consider multiple controllers)

 

1c) Multiple Spindles on the Client: Your Operating System functions on your hard disk and is constantly accessing it. Whenever your OS needs to access the disk, it ties up the disk for that moment because there is only one head and one spindle in the disk. Multiple partitions on a single disk don't change this and thus don't make any difference. When IronCAD loads or saves a file, not only is it reading/writing that file, it's also working with temporary files. So if you only have one disk IronCAD could be competing for usage of that disks resources with MANY other processes occurring simultaneously. If you increase the number of spindles and heads that your OS can work with, then you decrease the time any single process has to wait to access the disk. With this in mind an ideal setup for optimum disk access rates would be:

 

Disk1) Operating System and IronCAD Only, no temp or page file

Disk2) Scratch Disk for temp files only

Disk3) Page disk for the page file (Better to have enough ram so that your OS NEVER has to page though)

Disk4) Data disk that you store your files on

With this setup you're essentially multitasking.

 

[Guest ebedding]

Chris,

 

Thanks for the detailed reply. It sounds like lots of RAM and 3-4 hard drives to divide the workload at the workstation level is a good starting point. Would multiple lower cost 7200RPM HDD's still be helpful?

 

1. Is the new 1066Mhz RDRAM significantly faster than the old PC133Mhz RAM? Is 1G RAM needed for large assemblies?

 

2. Is an L2 transfer cache of 512KB helpful for increased IronCad speed? (Now Available on the Pentium 4)

 

 

Thanks again for all the info.

 

 

 

[Robert Andersson]

You may also want to look in to using Teamvault.

It have some really nice features.

You place all documents on a local Teamvault webserver

All clients have local mirror folders.

Then if you check out a large assembly from Teamvault, the system checks

if you already have some of those files locally in the mirror folders. Is so they are used.

All this is fully automatic, and the user don't have to think about it.

 

Also you can check out the main assembly with out locking all sub assemblys up, which you cant do without Teamvault.

Then if you click on a sub assembly that you don't have checked out, Teamvault asks you if you want to check that assembly out now.

 

Robert Andersson

[Chris Lohman]

Having multiple ata drives should help a bit due to the fact that you have multiple spindles but I personally wouldn't consider ATA for a CAD workstation if speed is the ultimate goal. Simply having enough ram for IronCAD to uncompress and load the file into ram instead of having write to page/temp is even a good place to start. ATA today is the "consumer" disk that is cheap for a reason. It's difficult to say how much you would benefit from SCSI vs. ATA for many reasons, one of them being I don't know anything about the files you're working with or the systems you currently have. I can only make general statements such as: If your goal is speed, SCSI is the speediest. Are the ATA drives enough if you divide the spindle load; I just can't say since your idea of "good enough" is most likely different than mine.

 

The transfer cache improves processor operations but I wouldn't expect to see it have much of an impact on load/save. It would certainly help in computational arenas but on load/save the main factors are disk access speed and ram.

 

I came up with some interesting numbers this morning that I thought might help the community.

 

I have a very complex assembly with a couple of hundred linked scenes that before 5.2 just would not load at all. It was soo complex that IronCAD would crash. I just tested this file in 5.2 on two seperate workstations and have some metrics. The assembly size comes out to approx. 60 megs.

 

System1

P3/500 1.5 gigs of ram

Single 10k rpm 9 gig scsi disk

Load time = 20 minutes

 

System2

P4/2.2 Xeon 2 gigs of ram

Single 15k rpm 18 gig scsi disk

Load time = 3 minutes

 

 

 

That's a very good point about Teamvault actually that I hadn't considered: When you're using the filesystem only, as soon as you load a master scene all of the linked in scenes are locked and loaded every time. If you're using Teamvault it only loads what you absolutely need so it seems faster because it's smart enough to only do as much work as is necessary.