A4091 performance

[Damion]

Recently picked up a 4091 for my A4K, and was wondering what type of performance this card is capable of. I know it's 10 MB/s in theory, but has anyone achieved anywhere near that?

Here's the full story (clipped from my similar thread over at EAB):

Hello,

Can anyone confirm a working fast synchronous mode on their 4091?

I recently found a DKB Rev B board (40.13 ROM). Synchronous mode yields ~4.4 MB/s, 2.6 MB/s asynchronous. These exact drives do just under 10 MB/s on my TekMagic 2060 (same NCR controller).

The "fast sync" mode, which should effectively double the raw transfer rate, doesn't seem to work - the dipswitch on the rear has no effect whatsoever.

Hardware is an A4000 with CSMK2 '060.

My thoughts so far:

1. Maybe the fast sync mode never worked on the 4091
2. The ROM code has some problems with the '060

Apparently, 9 MB/s or so is doable on the similar 4000T scsi, so I'm wondering if the 4091 ROM was simply never finished. The 4000T at least seemed to have several updates to its scsi device over the years.

Any ideas? Does anyone happen to have the older ROM code (40.9) laying around?

Otherwise, the card seems to work great... just the lack of a working fast sync mode.

[Damion]

Interesting issue. I wonder if you could use the 4000T SCSI patches on the 4091 driver...

Just for kicks, I tried installing a more recent version of the 4000T NCRscsi.device in the flashrom of my DENEB. It shows up as resident, but doesn't override the ROM that's physically on the 4091.

I think I'll fire up REMUS and burn a 27C256 for the 4091. I'd be _shocked_ if it worked, but I suppose it's worth a shot...

[Damion]

Well, looks like there aren't any patches for the 4091. I tried burning a ROM for the 4091 with the early 4000T NCRscsi.device, with the predictable result of nothing happening. Looks like 4.5 MB/s is the best it's capable of, at least with my '060 card.

The warpscsi.device from the samecalled turboboard connected to a uwscsi hd achieved 9.2 mb/s on my 68040 a4000. Dont have experience with 4091 though...

That's about what I get on my TekMagic, and was hoping for similar performance with the 4091. (Bought the 4091 after buying a Fastlane, which sadly arrived broken.) I'd stick with my DENEB, which works great... unless I attempt copying files between devices on the DENEB, then I get constant read errors.

Anyone have an old 4091 40.9 ROM they wouldn't mind parting with?

[Damion]

Thanks for the replies!

I figured it might be unrealistic to expect similar performance as accelerator SCSI via the Zorro bus. The Fastlane is reported to do up to 8 MB/s, but I believe it also eats more CPU. I've also found some info buried in the usenet archive indicating that the 4091 software wasn't fully optimized before C= tanked, and fast synchronous mode wasn't yet working. :-o Anyhow, filesystem operations are just as fast as my TekMagic, and RSCP indicates it leaves just over 90% CPU free during the benchmark - compared to 98% on the TekMagic, so again, not too bad for Z3 DMA.

Tj - sending you a PM buddy, thanks for the offer. :pint:

[Damion]

Just a few final notes on this (and a big ty to Tj for sending me the 40.9 ROM!). I found some info buried in the usenet archive from Dave Haynie, which essentially mirrored what was written above by SpeedGeek. If the 4091 was able to do anywhere near 10 MB/s over Z3, it would completely hog the bus and lock out the rest of the system from doing anything else. This is exactly what the Fastlane does to achieve its 7-8MB/s, as I found out while comparing the two controllers with RSCP.

A snippet from Haynie's usenet post:

The A4091 tries not to saturate the bus for extended periods
of time. In theory, the A4091 could take the bus and hold it for
seconds with a really fast drive. That would cause real problems for
any programs, serial bus interrupts, etc. So the A4091 tries to be
nice about things and gives the CPU some time on the bus, even when
SCSI can go about as fast as the Zorro III bus.

edit: OK, the Fastlane hogs the CPU in Buster 9 mode, some 20% free at 5MB/s and none free at 6-8MB/s. In Buster 11 mode, transfer rate and free CPU appear roughly equivalent to the 4091. Unfortunately, Buster 11 mode is unreliable here and hangs my system(s) after a minute or two.

Seems the solid 5MB/s of the 4091, while leaving over 90% CPU free, isn't so bad after all. :pint:

[Damion]

It's been a while, but I remember 4-4.5 MB/s with very little CPU use. Filesystem operations with PFS3 were really quick, too, not far off from my TekMagic or even a little better.

[Damion]

Hi Dan,

Does the bus speed mean that much to what you do?  Could you use a UWSCSI controller instead?  A USB external drive gets 5 MB/s on a Deneb, would that help you out, or a FastATA card with one of the upper range PIO modes?

Actually, the performance of the 4091 is more than enough for what I need to do (casual tinkering, games, demos, etc). Filesystem operations are just as zippy as my accelerators with SCSI, so the system feels similarly responsive. And the CPU use is negligible compared to IDE, or even the Fastlane. The real challenge here was simply figuring out whether or not the 4091 was performing at its best, which (tinkering with the hardware) accounts for a majority of the fun for me. An MK3 with UWSCSI would be ideal, but maybe too easy... LOL!

I have a DENEB installed, it's a nice, useful piece of hardware. Unfortunately I had one major problem - read errors when transferring files between devices on the DENEB. For example, copying MP3s between a flashdrive and the internal storage inevitably resulted in an unrecoverable failure. I was unable to resolve this after many countless hours trying, and could replicate the fault on a number of different systems (2000, 3000 and 4000) and hardware/software configurations. Copying between accelerator SCSI (or internal IDE) and the DENEB worked fine. Accelerator SCSI wasn't an option on my A4K/CSMK2, since it's clocked at 66MHz.

Also - CPU usage with the DENEB in DMA mode is similar to the Fastlane; it sacrifices quite a bit in that department for the sake of a higher transfer rate. This also explains why the DENEB works in DMA mode with Buster 9 in conjunction with the Fastlane and its Buster 9 "workaround" - there really isn't the same kind of efficient DMA going on like you have with the 4091 or accelerator SCSI... But hey, it's potentially faster, more compatible, and at least it works!

I have to say I'm 100% satisfied with the current setup, MK2 Cyberstorm@66MHz, DENEB + Melody 1200 Pro, A4066 ethernet, and the 4091. (The 4091 has an AztecMonster SCSI ->CF adapter and 16GB SanDisk onboard.) I've successfully transferred a ton of files between the 4091 and DENEB, and had the system up for a few days now with no crashes or resets. Aside from easy USB access, flashrom for custom kickstart+modules, and providing a clockport for the Melody, the DENEB also serves an additional useful purpose: without it, the 4091 refuses to see attached drives so long as the ethernet board is installed. Plop in the DENEB, and everything works just fine. :pint:

You should do some tests of the A3000 SCSI.

Incidentally, I did a bit of that this evening, with both the stock '030 and a few different accelerators in the mix. I get 3.8 MB/s from an old Quantum Fireball with some 80% CPU free. The Fastlane in the same setup will do 7 MB/s (with a fast enough drive), but it eats all available CPU time. Buster 11 mode (which, btw, works reliably only on my A3000 sans accelerator) yields ~5 MB/s and still sucks massive CPU.

[Damion]

I guess with the Fastlane saturating the internal bus (including Z3 overhead from Buster) it just stalls the stock CPU.

Makes sense, the CPU is locked out rather than being utilized, as the SCSI controller is still doing all the work. So it's still technically DMA (if I have that correct).

An accelerator board with RAM might solve this problem.

When the Fastlane's Buster 11 mode works long enough to run RSCP, it does perform about as well as the 4091, even a touch better. Problem is it locks up on both the 3000 and 4000, with several different accelerators, unless the datacache is disabled - which then cuts CPU availability almost back the levels of Buster 9 mode. Regardless, 7 MB/s on a stock 3000 over Zorro isn't too shabby, IMHO.