PDA

View Full Version : B&W with dual 15k Cheetahs



goliveira
10-25-2000, 07:11 PM
I am running these dual 15k Cheetahs in a B&W ( yes the troublesome one) with a 400 MHz G3. My scores are not that greater than it the 9600 which makes me happy.

Reads were 69MB a sec sustained
Writes were 40 Mb a sec sustained.

Running of a Miles U2W with firmware 1.06 and Apple/Foxconn cables.

My reads in the 9600 were 69 and writes @ 36.

The write scores do not change from a single drive or raid setup. Is the Miles 2 is the bottle neck here. Maybe the drive write speed would increase with the Atto card.

George

[This message has been edited by goliveira (edited 25 October 2000).]

[This message has been edited by goliveira (edited 25 October 2000).]

kaye
10-25-2000, 09:09 PM
George,

My PTP with XLR8 500MHz G3 in the Power Computing Step-by-Step forum:
Test 5 - one X15 and one Miles2
Test 6 - two X15 and one Miles2, striped
Test 7 - two X15 and two Miles2, striped

ATTO EPT 8MB, 2 Cache Segments, no Disk Cache, in MB/s:
Test 5 - SR 40.33, SW 40.20
Test 6 - SR 70.12, SW 40.39
Test 7 - SR 77.73, SW 75.50, PR 133.02, PW 76.00

SpeedTools QuickBench:
Test 7 configuration:
Xfer Size---Sequential Read---Sequential Write---Random Read---Random Write
1 MByte------75.919 MB/sec-----70.915 MB/sec----53.467 MB/sec--60.527 MB/sec

This is how you break thru the 40MB/s SW barrier in a vintage Mac with two Bandit chips. Two or more X15's or 10k Cheetahs on two Miles2. To exceed SW 40MB/s, takes two Miles2, bus duplexing. It is the PCI bus architecture of a vintage Mac that limits SW performance. I would think that in a B&W with a single Miles2 and two X15's you could break that limit. That is puzzling. The ATTO card dual-channel UL3D certainly should break that limit with one X15 on each channel. In a vintage Mac, the UL3D is slower than a single Miles2, again bus archtecture. My 2¢ k

Ton
10-26-2000, 02:20 AM
More or less same thing here:
4 x 9g 10k cheetah's / 1 Miles 2: 40mb/sec SW.
4 x 9g 10k cheetah's / 2x Miles 2: 61,5mb/sec SW.
(PM 9500/G3 400)
Performance gain in SR much less dramatic.
Ton

goliveira
10-26-2000, 07:48 AM
Thanks for the replies. Since I can't boot of a raid in this machine, I will keep them as seperate drives and enjoy the 40 MB write off each.

K.
Its weird how the Miles 2 maxes out @ 40 MB write. Thats the only explaination. The only performance boost by adding the second 15k Cheetah were the reads. On an Ultra 3 board, these Cheetahs should be reading 80 MB a sec. Rob @ barefeats confirms this with his review. The Adaptec 39160 ( with 3 15k CHeetahs) had pretty much matching read and writes @ 110 MB a sec. He said with the Atto Ultra3, reads were 110 but the writes fell to 89 Mb a sec.
http://www.barefeats.com/hotrod2.htm

Hope Atto's tech guys read that. Don't know much about their tech support but I got a feeling they are not like Initio. Maybe Magician can send them an email with that link.

George

kaye
10-26-2000, 10:33 AM
George, Ton, and magician,

I still think that there is something wrong with your striped RAID, 1x Miles2, 2x X15's in the B&W 400MHz G3. Have never tried a Miles2 in a machine later than a vintage Mac however.

magician, what do you think? Is the Miles2 limiting him to 40MB/s sustained write? Certainly not bus architecture? BTW, magician, would you post what PCI chip corresponds to later Macs?

Beige G3?
B&W?
etc.?

Below I have listed my last tests of the X15's from the Power Computing Step-by-Step forum:

Test 7 - two X15 and two Miles2, striped, one X15 per Miles2
(SoftRAID driver set for 256 Maximum Concurrent IOs and 991 Memory Pool Size, then X15's set for 23 Cache Segments via FWB Configure 3.0.2, in that order)
Test 8 - four X15 and two Miles2, striped, two X15 per Miles2
(SoftRAID driver set for 768 Maximum Concurrent IOs and 1369 Memory Pool Size, then X15's set for 23 Cache Segments via FWB Configure 3.0.2, in that order)

ATTO EPT 8MB, 2 Cache Segments, no Disk Cache, in MB/s:
Test 7 - SR 77.73, SW 75.50, PR 133.02, PW 76.00
Test 8 - SR128.60, SW 75.78, PR 137.13, PW 76.27

SpeedTools QuickBench:
Test 7 configuration:
Xfer Size---Sequential Read---Sequential Write---Random Read---Random Write
1 MByte------75.919 MB/sec-----70.915 MB/sec----53.467 MB/sec--60.527 MB/sec
Test 8 configuration:
1 MByte-----128.808 MB/sec-----71.736 MB/sec----81.791 MB/sec--68.745 MB/sec

MacBench 5:
(Note, these X15's are brand new and, IME, they get faster as they break in)
Test 7 - Disk 4212, Pub Disk 3440 (only ran this test once)
Test 8 - Disk 4503, Pub Disk 3622 (1st test)
Test 8 - Disk 4536, Pub Disk 3659 (2nd test)
Test 8 - Disk 4546, Pub Disk 3666 (3rd test)

Up to now, all X15 tests have been conducted using the Initio twisted-pair LVD cables and Initio LVD terminator so that I could assess their capabilities. In Test 8, I experienced a few hangs in MacBench 5 Disk and Pub Disk (BTW, I think those tests are an excellent exercise of your HD configuration). I initially thought the problem might be with the XLR8 G3 overclocked, so I backed off from 540/270 to default 500/250. The problem continued. My assessment is that the Initio cables and terminators have surpassed their limit with two X15's on each Miles2 or perhaps even one X15 on each Miles2, since I only did one MB5 test in that configuration. I am going to go back and recable the X15's with my Gurus Granite LVD cables and LVD terminators.

My conclusions, two Miles2, each driving one striped X15 (one on each Miles2), are a worthwhile investment for vintage Macs that have two PCI BusMaster Bandit chip slots. The gains over using one Miles2 driving two striped X15's are very significant, nearly double in sustained writes. Two Miles2, each driving two X15's (two on each Miles2), is where it gets questionable for me. Yes, the sustained reads go up a whopping amount, but is the overall increase worth the extra $1000 for the two extra X15's? Long term probably not for me, depends on your assessment of the gains, and my thinking now is that eventually I will pull two of the X15's for a later project and stick with two Miles2 each driving one X15 and striped. It certainly is a thrill ride though with the four X15's. k

goliveira
10-26-2000, 10:55 AM
k,

Darin is doing just that. Intalling dual Miles2 with dual 15k Cheetahs on each. He is getting poor results and I sent him the thread. Hopefully he can get his scores near yours. Those are amasing k.

In my 9600, I ran dual 9 Gb Cheetahs with the 350 604ev. My reads writes were 36 MB a sec. After installing the xlr8 G3 500 MHz, they were at 35MB a sec. The dual 15k Cheetahs with the same setup stayed @ 36. Showed how either the bus or Miles2 were at the max throughput.

Taking it to the B&W increased the single scores to 40 MB writes which amases me. But cannot get them over 40 even adding a second 15k as a raid 0. Shows the Miles 2 is maxed out. The dual Miles2 should give you up to 80 if the maxed out theory is true. The dual 15k Cheetahs on a Ultra 3 board had reads and writes @ 80.

G.

goliveira
10-26-2000, 12:57 PM
where is M?

haven't seen any activity from him on the forums all day.

Hoping he is not on a plane to Montana http://macgurus.com/infopop/emoticons/icon_biggrin.gif

G.

Louie
10-26-2000, 01:31 PM
I was wondering the same thing.

magician
10-26-2000, 03:16 PM
no, I just went to a Montana of the mind, and finished reading Footfall by Niven.

ok...you got Grackle in beige G3....which explains its better performance compared to vintage PCI Power Macs....in other words, slightly better PCI performance in beige G3, even though it is still 32-bit PCI.

I'm not sure which PCI controller is used in Yosemite....someone needs to either look at the controller on their logic board or check out the Developer's docs. I'm thinking it is Grackle, even though it is 64-bit PCI.

in Sawtooth, you transition to UniN, aka UniNorth. It is native 64-bit throughout.

G is correct: the Miles2 is a 32-bit PCI card, and that would account for the performance ceiling he is seeing in the Yosemite.

K is correct when he notes that bus duplexing a pair of Miles2 in a dual-Bandit Tsunami breaks the 40MB/sec barrier. I think you will see that sustained write barrier whether you are using a single Miles2 in a Yosemite or in a Bandit machine....though I'm pretty sure we have beige G3 benchmarks on the FTP site (and Greg Santilli can jump in here, as well) that confirm that 32-bit PCI Miles2 delivers slightly better sustained writes on the Grackle machines. I would hesitate to estimate without the benchmarks in front of me.

as for Darin, you guys need to remember that he is apparently laboring beneath a significant karmic debt, presumably carried over from a previous lifetime as a molester of guinea fowl, and he will be the one among us who is doomed to address drives at SCSI ID7 as he thrashes around in his personal hell in the middle of the night, wrestling with video cards that no one else is running.

I'm thinking that one of the pitfalls we suffer as Mac users is that we are tempted to do all things on one machine....simply because we know that we can....but when you mix Formac boards and 3Dfx boards in already complex Tsunami subsystems, you are simply asking for trouble.

who supports the Tsunami anymore? No one, really. The Formac boards and 3Dfx are really engineered to work in new Sawtooth, or at least other New World machines.

that said, we all know that Tsunami remains the most flexible and adaptable architecture, and used conservatively, enables us to run certifiable behemoth systems. How many other guys do you know that are running three displays off their Macs? And who are also trying to bus duplex? And who are also running the fastest video cards built? Just us. And most of the time, assuming we're not addressing drives at ID7, we get away with it.

I agree with George: definitely something going on in video card firmware somewhere, but getting anyone to support configurations like these is probably hopeless. Remember that software engineers have to build with a notional system in mind: at max, that means dual displays, one SCSI card, etc.

The fact that this stuff works at all, in my opinion, is testament to Apple's brilliance in laying out the architecture. Drivers built to spec actually work pretty well.

in the meantime, we should all encourage brother Darin to quit running so much ridiculous stuff in his extensions folder, and move his gaming cards to a dedicated gaming machine. Use work machines for work, and play on another machine. Simplification, in short.

now, I gotta go work, you crazy guys.

this forum is nutso! I take one night off, and look at what happens!

http://macgurus.com/infopop/emoticons/icon_biggrin.gif



[This message has been edited by magician (edited 26 October 2000).]

goliveira
10-26-2000, 04:18 PM
Darin should get a playstation and keep the games off his main machine.

From a story I read, the Playstation can crunch numbers twice as fast as an Apple G4.

Their is a chip labeld bridge near the zif made by Digital. It says 66 MHz but no Grackle ot Uni.

I dunno.

G.

[This message has been edited by goliveira (edited 26 October 2000).]

magician
10-26-2000, 06:20 PM
the Digital chip is the bridge.

Yose may use UniN.

goliveira
11-28-2000, 12:02 PM
I installed an Atto dual channel U3L card with the dual x15 one per channel. Stripped with Softraid 2.2.2 and formatted HFS+.

SR 75MB /sec
SW 54MB/sec

The Atto card is in slot 3. The machine doesn't care what slot its in. I actually saw 3 MB a sec better results in slot 3 than in 1. Ran the test twice in both slots with the same results.

The write speed is disappointing. I truly hope something there is a jumper in the wrong place since the write speed should be greater. Will pull the drives and check them completely and retest. Will also test them in a Sawtooth since they have greater theoretical throughput.

g.

magician
11-28-2000, 04:42 PM
interesting....those results are very similar to what we see when testing in beige G3's.

have you implemented all firmware updates on that machine? Running OS9.04?

goliveira
11-28-2000, 04:52 PM
yep, latest everything including the Xlr8 firmware update.

I just ran Quickbench on the setup and here are the results. The numbers fluctuate abnormally. There is some issues here on my end.
<table summary="QuickBench 1.5 Test Results File" border="0" cellpadding="2" rules="none" align="center">
<TR><TD>Xfer Size<TD>Sequential Read<TD>Sequential Write<TD>Random Read<TD>Random Write<TD>

<tr><TD>1 KByte<TD>504.439 KB/sec<TD>6.142 MB/sec<TD>321.564 KB/sec<TD>6.550 MB/sec <TD>

<TR><TD>2 KBytes<TD>8.433 MB/sec<TD>9.332 MB/sec<TD>1.685 MB/sec<TD>12.782 MB/sec <TD>

<TR><TD>4 KBytes<TD>7.105 MB/sec<TD>19.679 MB/sec<TD>2.949 MB/sec<TD>21.241 MB/sec <TD>

<TR><TD>8 KBytes<TD>39.397 MB/sec<TD>33.762 MB/sec<TD>4.659 MB/sec<TD>32.593 MB/sec <TD>

<TR><TD>16 KBytes<TD>30.287 MB/sec<TD>44.051 MB/sec<TD>4.694 MB/sec<TD>2.591 MB/sec <TD>

<TR><TD>32 KBytes<TD>57.689 MB/sec<TD>53.907 MB/sec<TD>8.961 MB/sec<TD>68.591 MB/sec <TD>

<TR><TD>64 KBytes<TD>75.510 MB/sec<TD>17.850 MB/sec<TD>17.520 MB/sec<TD>4.880 MB/sec <TD>

<TR><TD>128 KBytes<TD>59.987 MB/sec<TD>38.720 MB/sec<TD>27.185 MB/sec<TD>24.928 MB/sec <TD>

<TR><TD>256 KBytes<TD>54.940 MB/sec<TD>45.278 MB/sec<TD>32.185 MB/sec<TD>36.629 MB/sec <TD>

<TR><TD>512 KBytes<TD>71.282 MB/sec<TD>43.454 MB/sec<TD>47.912 MB/sec<TD>38.552 MB/sec <TD>

<TR><TD>1 MByte<TD>76.903 MB/sec<TD>44.687 MB/sec<TD>53.180 MB/sec<TD>42.987 MB/sec <TD>


</table>

g.

[This message has been edited by goliveira (edited 28 November 2000).]

goliveira
11-28-2000, 05:55 PM
I ran Quickbench with extensions disabled and I got:

76 Mb/sec read
71 Mb/sec write

Going to give this machine the works this week and hopefully won't be looking for an extension conflict.

g.

magician
11-28-2000, 06:23 PM
now that looks better to me.

disable ACTION menus.

(just kidding--I know you aren't running any of that garbage).

http://macgurus.com/infopop/emoticons/icon_biggrin.gif

goliveira
11-28-2000, 07:53 PM
I have no 3rd party stuff in my system folder. Since I am still testing this machine thoroughly.

I am no Darin who has an inventory of extensions. Doesn't he own a copy of every 3rd party utility and every game? Hopefully he has then setup in sets. http://macgurus.com/infopop/emoticons/icon_biggrin.gif

The computer guy how introduced me to the Mac back in the Quadra days, setup one computer. Usually my boss how uses a crap load of programs. He would then copy this system folder to every Mac in the building! Clean up was a mess. Had 7200s running with over 175 extensions and over 100 fonts in the font folder with Appletalk and filesharing enabled!!!

My extension folder stays usually under 100 items. Now with this ATI board, its closer to 120. Hate them.

g.

magician
11-28-2000, 08:44 PM
I'm in touch with that emotion.

we didn't sell ATI for a long, long time...in fact, we just started selling their boards with the release of the Rage 128's.

before that, we sold ixM and Formac (aka MacTell).

then ixM went under, Formac yanked their boards from MacTell, then MacTell went under, then ATI wooed us with the secret demo of their ATI boards at MacWorld-SF a couple of years ago. Little did we realize that they wouldn't release the damned things until the following July, I think it was.

they do use a shit-load of extensions. I always hated that.

PENDRAGON18
11-29-2000, 11:25 AM
The chip on the YOSEMITE is a GRACKLE-4, on my REV2.

I wouldnt think that the MilesU2W would max out at 40MB/s on the YOSEMITE. Just because the YOSE has 64bit/33Mhz and one 66Mhz/32bit PCI slot doesnt mean the SCSI cards needs to take advantage of this. PCI 1x is 33Mhz/32bit and it should (in theory) handle up to 133MB/s sustained throughput. So even if its not at full efficiency, 80 or 70MB/s should be achievable. I wounder if the 66Mhz/32bit PCI would work best for the MilesU2W?

Or could it even take advantage of the faster clock?

One thing I dont like about these benchmarks. Most new drives have a large cache built in. So I really think the test data size should be 8MB, 16MB and 32MB. Otherwise your not really running off the drive, but the drive's cache.

------------------
Have fun storming the castle!

goliveira
11-29-2000, 12:01 PM
P,

the Miles U2W maxes out @ 40 Mb write. It can't go any faster. No matter what bus your using.

M,

I ran the setup of dual x15 with the Atto card in a Sawtooth with these results:
75 MB / sec read
82MB /sec write

The Yosemite is staying around 54 Mb /sec write even with a 500 Mhz G3.

Using Quickbench, I got a test confirmation of 71Mb / sec write ( extensions off, dual channel setup) but that was at 32k, it fell to the 54 MB level right after that. This leads me to believe it's bus issue on this Yosemite Rev 1. Guess I'll get a new machine @ Macworld NY when they release the G4e with the rev2 PC 133 etc.

Will do some more research. Maybe I should try different firmware versions. See if 1.5 works better.

g.

magician
11-29-2000, 02:44 PM
you weren't able to massage extensions to get better thruput from the Yose?

at least we know your hardware isn't defective or misconfigured....

what cabling and termination are you using?

goliveira
11-29-2000, 03:06 PM
2- Apple Foxconn cables from the liquidation page. I tried each cable in a single channel dual drive array with same results from each channel. It also pumps data through in the Sawtooth which doesn't seem like a hardware issue though I will order the Granite stuff after the holidays.

With extensions off or booting of a cd yields similar results.

This machine has a Grackle 4 bridge chip. Same as the Rev 2.


g.

[This message has been edited by goliveira (edited 29 November 2000).]

magician
11-29-2000, 05:17 PM
I think it's a Grackle vs UNI-N thing.

Santilli
11-29-2000, 08:46 PM
Hi
Just got done installing 4 X 15's in a beige minitower, 333mhz. 384 MB of ram.

Results with ATTO Express Protools 1.2:
Max transfer size is 8
Sample Size is 8
Cache off:
Peak Reads:
74.60 MB/sec(all numbers following are in MB/sec)
SR
70.31

Peak Writes
65.28
Sustained writes
64.72

Cache on is almost identical.


HI
The following are from a 3 x X15(ST318451) Raid 0 using an ATTO ultra 3 dual channel card. Two drives on
one channel, one on the other.

Machine is G3/333 minitower.

Max Transfer rate 8MB
Sample Size 8
All tests done on empty drives
Cache on:

Peak Read 75.33 MB/sec(all figures MB/sec from here out)
Sustained Read 57.83
Peak Write 61.53
Sustained Write 55.42

Cache off
Peak Read 74.48
Sustained Read 57.80
Peak write 62.08
Sustained Write 55.42

Using SoftRaid 2.2.1

The mac is back!!!
GS

PS These are prior posts and discussions on the same subject.


Name
Date
Forum/Archive
Why Super Fast RAID?
Santilli
19 July 2000
RAID
Beige X15 Raid? How Many drives?
Santilli
18 July 2000
RAID
Back in the RAID game!
Santilli
23 July 2000
RAID
S900/Miles2/SoftRAID/3x Cheetahs Works
Santilli
26 July 2000
RAID
S900/Miles2/SoftRAID/2x striped X15 Cheetahs
Santilli
28 August 2000
RAID
Mfg. of a RAID Controller?
Santilli
01 September 2000
RAID
RAID... SCSI vs IDE
Santilli
04 September 2000
RAID
PC Raid cards: Why so expensive?
Santilli
18 September 2000
RAID
Raid 0-X15X3 G3/333 Benchmarks
Santilli
23 September 2000
RAID
Performance on Single channel RAID and launching apps
Santilli
25 September 2000
RAID
4 X X15 Raid 0 Benchmarks
Santilli
03 October 2000
RAID
2 or 4 drives in Raid 0 startup?
Santilli
10 October 2000
RAID
IDE RAID question
Santilli
11 November 2000
RAID

This didn't copy well, but the links above rediscuss, and provide facts relevant to this discussion.

Not only did I hit a wall with 4 X 15's, but adding a fifth X 15 actually slowed the system down.

Magician, did this help http://macgurus.com/infopop/emoticons/icon_biggrin.gif?

magician
11-29-2000, 11:25 PM
yeah, man.

good work, greg.

looks like Grackle maxes out around 65-75MB/sec. Interesting compared to single Bandit, which maxes out at 40MB/sec sustained writes.

UNI-N is definitely fastest at this point.

Santilli
12-04-2000, 12:30 PM
Hi M

Have you figured out the max throughput on the B&W chipset, and the G4 duals yet?

What chipsets do they use?

Thanks

gs

magician
12-04-2000, 02:37 PM
B&W uses Grackle. G4 dual uses UNI-N.

looks like Jorge has figured this out, with Grackle apparently maxing out around 75MB/sec or so sustained writes. Uni-N will support sustained writes around 140MB/sec using six or more Cheetahs. It may go faster. We haven't tested with more than eight Cheetahs.