64-bit/66MHz SCSI Goodness is GOOD FOR YOU!!!

[MichaelD]

12 Nov edit

Well, 2.66GHz was not P95 stable. If it can't run P95tt for at least 12 hours (preferably more) it's not a stable overclock.

I lowered the multi to 15 so I could keep my 166FSB and keep the FSB/Mem ratio at 1:1

I now have a P95 STABLE CPU speed of 2.50GHz. That's 17 hours, 46 minutes of P95TT stable. :beer:

A 900MHz overclock. Not bad.


















____________________________________________________________________________________
10 Nov edit

Thanks to my partner in crime, Marlin1975 (he turned me on to the 166 secret) I now have a full
1GHz STABLE overclock!!! Running 1:1 Mem/FSB ratio. System feels a little faster. Changed memory settings in the bios from Auto to DDR333. Changed jumper setting/bootup procedure to 166MHz. (Thanks, Marlin!)

/game show music plays, crowd screams, etc


NO! You LIE like a rug!!! <--you say

No! I speak the truth! <--I say

Here's the proof.


Da proof

Please note the CPU Specification of 1.60GHz, then note the actual CPU speed/FSB speed, etc. I'm showing both CPUs just for clarity's sake.

These LV Xeons are a dream. Granted, the TOTAL price of entry is steep (dual Xeon mobo, EPS PS, Xeon heatsinks, etc) but the overall system price is just a joke compared to building a comparable non-overclocked rig from scratch.

Two, 2.60GHz Xeons for $120. Not even on an exceptional day on EBay will you find that. :beer:

CPU based benchmarks have gone up a bit. Memory benchmarks are for the most part, identical. So are videocard benchmarks. (I need an X800P/6800GT!!!)

Thought I'd let you know.















______________________________________________________________________________________
Finally, I have my SCSI HBA in a proper 64-bit/66MHz slot. It's currently living in my dual-Xeon rig. (Asus NCCH-DL motherboard). Seat of the pants testing is one thing, but benchmarks are another.

These benchmarks were made with the exact same setup; same partition size, file system (NTFS), etc. The only thing that changed is the slot.

Here's the old benchmark; card in a standard PCI slot (32-bit/33MHz)



Here's the new benchmark; card in a 64-bit/66MHz slot.


Feel free to open the windows side by side. I did. Repeatedly. :beer:

17MB/s increase on writes
30MB/s increase on reads

Ironically, SANDRA shows my transfer rates to 7MB/s writes and 4MB/s reads. :roll: Bahahaha! I wish I had a FOUR drive stripe. :evil:

Sharkkeeper will be along shortly with his Drool Screen (as I call it.) A screenie of results from like a 20-drive, fiber channel array...something like 600MB/s reads. :eek;

 

[imported_michaelpatrick33]

MMMM SCSI and 64bit and 66mghrtz MMMMM.

 

[Pariah]

What's wrong with the first benchmark? 101MB/s is poor even by 32/33 standards. You should be getting around 120MB/s. The write benchmarks are pretty awful for both, might want to look into that.

 

[MichaelD]

Originally posted by: Pariah
What's wrong with the first benchmark? 101MB/s is poor even by 32/33 standards. You should be getting around 120MB/s. The write benchmarks are pretty awful for both, might want to look into that.

I disagree. Consider for a moment that on a standard 32-bit/33MHz PCI slot motherboard that the PCI bus is shared with the IDE channels along with the network card, soundcard and whatever else is plugged into it, and suddenly 101MB/s is pretty darn good considering there's a theoretical max transfer rate of 133MB/s.

 

[Pariah]

Unless you were running one of those old broken VIA chipset motherboards, you should benchmark at least over 115MB/s with that setup. I can produce 120MB/s on 32/33 MHz and I've seen many sites do the same.

 

[MichaelD]

Originally posted by: Pariah
Unless you were running one of those old broken VIA chipset motherboards, you should benchmark at least over 115MB/s with that setup. I can produce 120MB/s on 32/33 MHz and I've seen many sites do the same.

It was an NForce2 mobo. Abit NF7-S. I'd like to see pics of your benchmarks, please. Same two, 15K drives, right?

 

[Pariah]

Doesn't matter what drives, and setup capable of over 120MB/a would hit the same limit.

 

[sunase]

I have a pair of 15k.3's on a bus limited controller that do 120 in RAID0 on a Gigabyte 8KNXP mobo no problem. I admit I have hardly any PCI cards taking bandwidth, though.

 

[sharkeeper]

Yes getting the most out of the PCI bus requires tuning your latency settings. I've seen 126MB/S before but you never see this out of the box unless you're lucky...

Those write speeds don't look so good. Try running the HBA with the following parameters:

Read Ahead: no read ahead
Cache: write through and standard i/o

Run ATTO again.

I've found that ATTO gets flaky when tested on arrays > 750MB/S. Must not like supersonic speeds!

Cheers!

 

[imported_Phil]

Originally posted by: sharkeeper
Yes getting the most out of the PCI bus requires tuning your latency settings. I've seen 126MB/S before but you never see this out of the box unless you're lucky...

Those write speeds don't look so good. Try running the HBA with the following parameters:

Read Ahead: no read ahead
Cache: write through and standard i/o

Run ATTO again.

I've found that ATTO gets flaky when tested on arrays > 750MB/S. Must not like supersonic speeds!

Cheers!

You are required, by law, to post screenshots of said 750Mb/sec benchmarks

 

[CraigRT]

I was going to do that with my Athlon MP box, Man, I should have just kept it and upgraded the CPU's.. LOL damn.

 

[MichaelD]

Originally posted by: sharkeeper
Yes getting the most out of the PCI bus requires tuning your latency settings. I've seen 126MB/S before but you never see this out of the box unless you're lucky...

Those write speeds don't look so good. Try running the HBA with the following parameters:

Read Ahead: no read ahead
Cache: write through and standard i/o

Run ATTO again.

I've found that ATTO gets flaky when tested on arrays > 750MB/S. Must not like supersonic speeds!

Cheers!

Thanks, Sharkeeper. I tried those settings and it completely killed performance. Writes dropped by 12MB/S and reads by almost 20MB/s. I quickly put it back.

As you pointed out during our conversation, the i875chipset is the problem here. It's not a real server chipset; just a really good chipset adapted for "server" use. Oh well. At least it's faster than it was before and I've got the dual-CPU goodness going on now.

 

[sharkeeper]

Writes dropped by 12MB/S and reads by almost 20MB/s

Something is wrong there. Those numbers should be in the 140's for reads and 130's for writes given a pair of MAS'.

1600's are very picky about the motherboard they are hitched to, however turning caching OFF should fix this problem. I always recommend running the cache ON for best "real world" results. Always.

Cheers!

 

[Markfw]

Why do I get 185 m/sec write and 196k/sec read on 5 10k drives in ATTO ? Is it that something is wrong, even though you only have 2 15k drives..... And mine went from 60/50k to those numbers simply by changing the slot, and we have almost the same SCSI card ?????

Sharkkeeper ??? Help ! Are mine in line ? Whats with Michaels ?? I would think he should be getting 120-160 ? Or are they not in raid0 ?

 

[MichaelD]

Originally posted by: Markfw900
Why do I get 185 m/sec write and 196k/sec read on 5 10k drives in ATTO ? Is it that something is wrong, even though you only have 2 15k drives..... And mine went from 60/50k to those numbers simply by changing the slot, and we have almost the same SCSI card ?????

Sharkkeeper ??? Help ! Are mine in line ? Whats with Michaels ?? I would think he should be getting 120-160 ? Or are they not in raid0 ?

As The Master explained to me, it's all about the chipset that controlls the PCI_X slots/raid function.

Apparently the Intel 63000 (something close to that...I've been up almost a full day now...) is more of an interim solution. Somewhere b/t a "ATX motherboard" and a "Server motherboard."

Whereas the AMD 8131 chip that controlls your PCI-X slot is more suited to a server, per se. Apparently, I've struck out on this particular part of my implementation.

That's OK though. I've got the massive Xeon overclock goodness going on. Where else has anyone found TWO 2.4GHz Xeons for $120?

for my first foray into The Big Leagues, I think I did pretty well.

 

[Markfw]

Thats OK, I was just worried for you. Good to hear you have an explanation. My next upgrade is to 4-6 73 gig 15k rpm SCSI Hitachi or Maxtor (whichever was the one that won the award) and an LSI Logic x320-2 controller. Don't expect to see this too soon, I may be well off, but I have a son in college, with no loans so far, and I want to keep it that way.

 

[TerryMathews]

I wouldn't give up hope just yet - the fact that the write speeds level off indicate that something's wrong.

What is the block size for your stripe?

Also, read http://support.microsoft.com/?kbid=332023 link - SCSI support in WinXP isn't exactly what I'd call robust. You may need to spend some time tinkering to fully realize your maximum performance.

Also, you're running at a funny FSB right? Have you tried POSTing at 200MHz FSB? I think you and I talked once about getting the 1.6LVs working on the 200FSB...

 

[Pariah]

Originally posted by: Markfw900
Why do I get 185 m/sec write and 196k/sec read on 5 10k drives in ATTO ? Is it that something is wrong, even though you only have 2 15k drives..... And mine went from 60/50k to those numbers simply by changing the slot, and we have almost the same SCSI card ?????

Sharkkeeper ??? Help ! Are mine in line ? Whats with Michaels ?? I would think he should be getting 120-160 ? Or are they not in raid0 ?

ATTO is not a good measure of performance. It does not generate consistent results. It has a habit of producing clearly wrong results often. There doesn't appear to be any pattern for what causes the odd results either. The only HD benchmark that I have seen generate consistent STR results for all configurations including RAID is Winbench's test.

 

[sharkeeper]

I've never had this "SCSI performance bug" surface on any of my systems. I don't use dynamic disks either. Of course none of my desktops have "dumb" controllers and this discussion revolves around intelligent controllers.

Anyways, WB99 does not use a write test and it isn't always accurate with RAID as well. HD Tach is generally useless as the user cannot specify a block read size and this is extremely important when gauging the sequential transfer speed of a striped array. This is why stripe sizes vary dramatically with HD Tach and people will adjust their stripe size for best HD Tach results which is NOT the way to tune their arrays.

If you really want to measure your RAID array performance under a condition it was designed for (a SERVER benchmark) use I/O meter.

With the 1600, those low speeds generally indicate a problem as severe as a conflict in your system. Funny thing is you will never (generally) notice it as STR has a very low impact on Windows desktop performance. Copy 20GB of raw data to that partition and you will notice, however.

Someone did mention an overclock. If you are running your FSB out of whack and your 64/66 slot is running at say 64/80, this could cause an issue as well. When troubleshooting any issues such as this it is imperative to run everything at default values before pointing fingers!

Why do I get 185 m/sec write and 196k/sec read on 5 10k drives in ATTO ? Is it that something is wrong, even though you only have 2 15k drives..... And mine went from 60/50k to those numbers simply by changing the slot, and we have almost the same SCSI card ?????

Those numbers are close if you are running with writeback cache. Turn it off to get raw r/w performance. Are you running a 0 or 5 stripe? It is very important to remember that if you use ATTO and benchmark with caching ON, the test runs in the cache and will produce identical results whether you have one drive or sixteen. The difference is when you start thrashing the hell out of with long transfers. The cache will fill up and average performance will drop to STR x n number of drives at 90% up to 250MB/S with that controller depending on your system configuration. With the cache off, the numbers scale correctly in the benchmarks. I recommend using the cache for best overall performance unless you find a specific need to have it off.

Cheers!

 

[MichaelD]

Hi guys,

Here are my settings:

Stripe size: 64KB
Adaptive Read Ahead
Write Back Cache
Cache flush=10 seconds

I'm overclocking and running a 150FSB. This doesn't have an effect b/c the NCCH-DL has a working AGP/PCI lock. CPU-Z shows AGP/PCI speeds of 66.6MHz and 33.3MHz exactly, so that's not it.

I cannot run a 200FSB b/c these particular Xeons can't take it. The way it works is that for a split second, the bios powers up the CPUs at DEFAULT multiplier (16x) times whatever FSB you've set. THEN pending a successful boot, it lowers/raises speeds to whatever you've set in the bios.

I cannot post at 12x200 which is the same 2.40GHz I'm running now b/c of the above stated problem.

16x200=3200MHz and my CPUs can't do it.

I have Ghosted my current install of WinXP. I may load up 2003 Server Enterprise just for kicks and see if there's a diff. WinXP/Server2003 have the same NT core...shouln't make a diff, should it?

I DL'd IOMETER; man what a HUGE program! Bit of a learning curve there. First time I tried to run it, it created a 5GB test file and filled up my OS partition. :rolleyes; Like I said...learning curve. I have to read the 48-page of documentation it comes with first.

 

[Markfw]

MichaelD, That card should be running in a 100 or 133 mhz slot. Or doesn;t that motherboard support those ? (mine has 3 busses, one maxes at 100, downgrades to 33 or 66, but everything on the bus goes down, the other is 133 mhz, goes down to 100 or 66 or 33 I think) If your bus really runs at 33, that will kill the performance I would guess.....

And make sure your other card (I think I saw another card) is not on the same bus. Actually I have three busses, as ONE slot is dedicated to legacy PCI 32-bit ONLY

 

[MichaelD]

Thanks Mark.

The NCCH-DL has two separate busses. A 32-bit, 33MHz bus with two regular PCI slots and a 64-bit, 66MHz PCI-X bus with two slots. The PCI and PCI-X busses are completely separate and don't interact.

Block diagram of the NCCH-DL

 

[sharkeeper]

The MegaRaid Elite 1600 is a 64bit 66MHz product. 100MHz and higher is PCI-X which that card is not.

Cheers!

 

[TerryMathews]

Originally posted by: MichaelD
I'm overclocking and running a 150FSB. This doesn't have an effect b/c the NCCH-DL has a working AGP/PCI lock. CPU-Z shows AGP/PCI speeds of 66.6MHz and 33.3MHz exactly, so that's not it.

I cannot run a 200FSB b/c these particular Xeons can't take it. The way it works is that for a split second, the bios powers up the CPUs at DEFAULT multiplier (16x) times whatever FSB you've set. THEN pending a successful boot, it lowers/raises speeds to whatever you've set in the bios.

I cannot post at 12x200 which is the same 2.40GHz I'm running now b/c of the above stated problem.

Did you swap around the CPUs? I've got one that can POST at 16x200 and one that can't. As long as the one that can is in CPU0, you're golden.

Also, did you make sure you were using the 1:1 memory setting and not the faster one? When you jumper and POST at 133FSB, you need to set the memory to 333 to get the proper 1:1 setting at 200. ASUS = stupid BIOS when it comes to that.

And yes, just because you've utilized a PCI lock doesn't mean it's not adversely affecting performance. Running the bus at 66.6MHz and the FSB at 150MHz means they're no longer synchronized - for large transfers the processor is going to have to wait for a 66.6MHz clock cycle to transmit data to/from the SCSI card. It'd be more of a latency issue than a transfer bandwidth one - but it still might be causing some other gremlins.

If you really want to check if it's a hardware problem or software one, download a Knoppix LiveCD that supports your SCSI card and try one of the Linux HD benchmarking programs.

 

[MichaelD]

Originally posted by: sharkeeper
The MegaRaid Elite 1600 is a 64bit 66MHz product. 100MHz and higher is PCI-X which that card is not.

Cheers!

Shiit. You're right.

/sharkeeper says "Yes, I know I am."

So, is this the answer to my "Why isn't my performance where I think it should be?" question?

Ahh, but the specs specifically say "66MHz." I know for a fact these slots are not 133MHz slots.