68K/early PPC with onboard AAUI 100mbps ethernet?

[jeremywork]

I learned a lot from this thread, but it looks like nobody ever posted test results... I have some here for those still curious:

Test file: Mac OS 9.2.1 Update.smi (84.4MB / 86,018,641 bytes)

Switch: TrendNET TEG-S24g Gigabit

Server: Power Macintosh 9600/1000(G4 [2MB Cache on board])

Apple PCI 10/100Mbps Ethernet card connected at 100Mbps

file hosted on RAM Disk over AppleTalk File Sharing

Client: Power Macintosh 950/100 (Quadra 950 with 100MHz Sonnet/Daystar 601 [1MB Cache on board])

AsantéFAST 10/100 Nubus connected at 100Mbps

AAUI connected via Apple Ethernet Twisted Pair Transceiver at 10Mbps

file downloaded to RAM Disk

Here's the cleanest test I could produce for now; I intended to test with the 68040 as well but for some reason the Quadra bombs on Memory extension if a RAM disk is enabled and the PPC card is disabled (or even removed, along with all the nubus. All combos of PRAMs and Reinstalls tried.) I give up on this for now. It's been many hours and countless restarts. 

None of the results are particularly thrilling, but YES the 10/100 card makes a noticeable difference.

 

[Gorgonops]

None of the results are particularly thrilling, but YES the 10/100 card makes a noticeable difference.

That's still amazingly slow though, wow. The classic OS does pretty much live up to its reputation.

(I used to netboot 486 and early Pentium machines into Linux over 10baseT networks and one machine could easily saturate the line. That best "100mb/s" score is only about half what 10baseT can do.)

 

[Cory5412]

It's definitely noticeable in the graph. Was it noticeable in actual usage? Did you feel like if you had one of those cards in your machine in the '90s it would've made a worthwhile difference?

Yeah.... none of those speeds is impressive for a 10-megabit link, let alone a 100-megabit link.

As your chart notes, 577 kilboytes is under half of the potential speed of a 10-megabit link.

One thing that would be interesting is seeing if upgrading the server to a newer OS and upgrading OT+AS on the client would have an impact. You'd move, in that circumstance, to AppleShare/IP and that might help. (Or, use ASIP6+ as the server instead of OS 9.)

 

[Cory5412]

Want to add: This is definitely interesting - in theory this is nothing the original hardware shouldn't be able to do.

In fact, it would be interesting to see more details about the OT+AS versions involved and whether or not upgrading them and maybe moving the server to more modern software (to force the IP link, where possible) would work.

Another test that would be interesting is, say, Netscape 3 or 4 download times. I don't know if it occurred in this thread but I eventually tried Netscape instead of IE on my old Macs and it does download files much faster than IE did.

I imagine the real use case of this card was if you were already using some kind of third party network stack that did better than this with onboard networking, or if you were doing multiples of something, I believe the AWGS6150, 8150, and 9150 might also have been mentioned.

 

[jeremywork]

I'm going to run a few more tests- I'll try from the Pismo's RAM disk over 100Mb from 9.2.2, and I'll also try using Netscape to download its own install package from system7today (this download yields a peak of 96Mbps on my Mac Pro, for the 12.1 MB file, so the test should be fair.) 

I also have several questions/observations others may have experience with:

The Quadra 950 has a partial implementation of Nubus 90 (enhanced rate is capable between cards on the bus, but not between the bus and the CPU.) I'm curious to know if the PDS slot in the 950 (which shares some of the address space on Nubus #5) is capable of acting as a Nubus 90 device. The only evidence I have to suggest this is the case is that my two SE IV cards are only even detected in the Quadra with it running off the PowerPC. Both updated first to 1.6.5 and then to 2.1b, both work great in the Quadra (up to about 8MB/s on an Ultra320 drive vs about 4MB/s on the built-in) but neither are detected by the IIfx in 7.6.1 or the Quadra running off its 68040. I'd love if my 840av would chime (it does once in a blue moon if you leave it humming long enough) because it has full Nubus 90 implementation (and 40 > 33) so we might see a better result there.

The second thing that piques my curiosity is the system clock timing. With the PPC Running, Sonnet Metronome reports a 100MHz 601, 33.3MHz CPU Bus & Cache, and 16.7MHz System Bus. Gauge Pro reports a 100MHz 601, 50MHz Backside Cache, 50MHz System Bus. Obviously both cannot be true, and Metronome reads the 68040 as a 67MHz CPU on a 33MHz bus, so its accuracy is limited on something this old. However, if the CPU is running a multiplier of 2 on an intermediary 50MHz bus, then running the real System Bus at 16.7MHz would make some sense, if not for the lack of performance. Gauge Pro shows 12.5MB/s Memory Performance (moving memory, 64-bit) when running from the PPC. Unfortunately, A ) it's PPC code, so I can't run it on the 68040 and B ) without the faster cache present, I'm not sure how the speed of the memory can be tested reliably.

The only dirty test I can think to try (because this machine refuses to have a RAM disk in 68k mode) is to pull the file down to the UW drive in both PPC and 68k mode several times to see if there's a significant throughput difference. 

If anyone had to guess- what's the most efficient OS to run on the Quadra to continue testing? I'll continue using 8.1 for the sake of consistency here. More tests to come.

 

[jeremywork]

It's definitely noticeable in the graph. Was it noticeable in actual usage? Did you feel like if you had one of those cards in your machine in the '90s it would've made a worthwhile difference?

Even after seeing this round of results I'm not inclined to remove the card, though if it turns out to make a much bigger difference in a machine with full Nubus 90 support, I probably wouldn't miss it too terribly in the 950. If the year was 1995 and my $7-10k worth of three-year-old computer could have 38% faster networking for a reasonable price, I would've considered it strongly.

 

[jeremywork]

Round 2:

(Connectix is enabled throughout until I specify otherwise)

Q950-PPC-8.1

Download http://download.system7today.com/officeupdates.sit (20.8MB; 21,912,064 bytes)

Destination: RAM Disk

Netscape Communicator 4.8

10Mbps

92 seconds / Avg 238 KBps / 1.91 Mbps

100Mbps

75 seconds / Avg 292 KBps / 2.34 Mbps

iCab and Internet Explorer both allow specifying the RAM disk as a download location, but according to the SCSI I/O it appears the browser caches the file to the disk first, then copies it to the RAM disk, making the results laughably slow and not worth mentioning for the purpose of this test. I'll do more testing with the UW drive as the destination later.

Now another round of testing with Mac OS 9.2.1 Update.smi, but served by the Pismo this time (the file took less than a minute to get from the 9600's ATA133 drive to the Pismo's RAM disk over 100Mb, so the 9600 really shouldn't have been the limiting factor.) I didn't mention, the 9600 runs 8.6. Pismo is 9.2.2, which has the option to 'Enable File Sharing clients to connect over TCP/IP' which I have enabled.

Source: Pismo's RAM Disk; Destination: Quadra's RAM Disk

100Mbps

143 seconds / Avg 601 KBps / 4.81 Mbps

As this is just a hair faster than the original test from the 9600 I'll just be testing with the 100Mb connection until I find something that produces a more dramatic change.

I then updated the AppleShare Client software on the Quadra from 3.7.4 to 3.8.3 and re-ran the test. (This update also allows 7.6.x and newer to authenticate OS X shares through Tiger, as well as Leopard but in read only.)

100Mbps

147 seconds / Avg 585 KBps / 4.68 Mbps was the best time I could get after several retries. (Most of these tests have been done more than once for sanity checks, the best times have been reported, as long as the rest have been close enough for margin of error.)

So that slowed down slightly...

Let's try something completely different- from the Pismo's RAM Disk directly to the UW SCSI drive attached to the ATTO Nubus Card.

100Mbps

138 seconds / 623 KBps / 4.99 Mbps

Well this certainly lends credibility to the partial Nubus 90 implementation theory, but these numbers are still so small it's hard to say with certainty. These numbers reproduce consistently from the Pismo's RAM drive, but from the 9600's RAM drive it doesn't seem to make a big difference whether the destination is the Quadra's RAM disk or its UW drive.

The last thing I'll do tonight is reboot the Pismo from 10.4.11 and mount its IDE-CF drive from the Quadra. For the first test I'll grab the file from the IDE and drop it on the Quadra's RAM drive.

100Mbps

188 Seconds / 458 KBps / 3.66 Mbps

And now from the Pismo's IDE to to the Quadra's UW drive.

100Mbps

203 Seconds / 424 KBps / 3.39 Mbps

And since this is a probable current day use case, I decided to run this test again with the 10Mbps connection (IDE -> UW): 

10Mbps

119 Seconds / 723 KBps / 5.78 Mbps

...

What?! So much for the last thing I do tonight...

AppleShare update must've improved the efficiency of AAUI, perhaps with a new driver. (Using the RAM Disk as a destination on the Quadra is still slower than the UW drive: 137 Seconds)

Pismo back to 9.2.2, keeping the file hosted on the RAM Disk, pulling it down to the Quadra's UW:

10Mbps

130 seconds / 662 KBps / 5.29 Mbps

[100Mbps]

[138 seconds / 623 KBps / 4.99 Mbps] [copied from above for reference]

So OS X shares are faster than OS 9 shares (IDE share from OS 9 is slower than RAM disk share, when downloading to the Quadra's UW.)

...and clearly I need a better driver for the 10/100 card. I'll be looking around for that.

Now I'll repeat the two tests from Netscape and call it a night; same file as above.

10Mbps -> RAM Disk

74 seconds / 296 KBps / 2.37 Mbps

10Mbps -> UW Drive

75 seconds / 292 KBps / 2.34 Mbps

100Mbps -> RAM Disk

70 seconds / 313 KBps / 2.50 Mbps

100Mbps -> UW Drive

70 seconds / 313 KBps / 2.50 Mbps

Both interfaces show improvement since the AppleShare update. The 10/100 card retains a small advantage here, though if no better driver is available I might rather have the free slot based on the AppleTalk results   This card may be designed for a machine with full Nubus 90 implementation. Hopefully it's just a driver issue though.

Anyone running 7.6.1-8.1 should install the AppleShare Client 3.8.3 update!

Future ideas:

I[SIZE=1.4rem]nstall AppleShare IP 6.1.1 on the 9600 if I can obtain an installer?[/SIZE]

Upgrade the Q950 to 8.6?

 

[Cory5412]

Even after seeing this round of results I'm not inclined to remove the card,

That wasn't, strictly speaking, my question, but fair enough. If I had one, I probably wouldn't pull it out either, although I'm not inclined to believe that the results we're seeing here are "meaningful" per se.

Yeah, I imagine the use case here really was if you had some kind of non-Mac server (IPX or NT) where appleshare service performance was noticeably better than with a Mac talking to another Mac. Like your OS X results. Netware/IPX was cited from time to time in old MacWorld for this reason.

The other thing I imagine this card being good for is if you had an AWGS 8150 or 9150 with some reasonably fast SCSI disks (or, just, several disks) such that you need to have multiple interfaces or such that upgrading to a switched network and a 10/100 port on the server will make "a difference."

As far as ASIP6 goes, there's an installer for 6.3.3 on Mac Garden. If you have a G3 or G4 of some kind around, that's maybe a better machine to run that on, since one of the options is a QS'02 9.2.2 restore image. You can mount the restore image and copy the whole thing to your hard disk, then pop in a serial number, and you're mostly ready to go.

It's a bit of an unrealistic test because that software is from 2002, a bit after the useful professional performance-sensitive service life of anything with NuBus in it, but it could be interesting nevertheless.

 

[Gorgonops]

This card may be designed for a machine with full Nubus 90 implementation.

The closest Apple got to a machine that meets that description is the faster versions (100mhz) of the 8100; the Nubus ASIC used on the other x100 PPCs is functionally identical to the MUNI chip that shipped with the Quadra 840AV; the difference is laid out in the "Enhanced Power Macintosh" Dev Note, but it seems to boil down that it has a somewhat better mechanism than the older chip for handling block moves, which was the major feature introduced in MUNI. Technically none of them are a "full implementation" because they all still only support a 10mhz base clock for transfers to and from the host; the higher 20mhz clock is still only applicable between cards.

Given the fact the highest score you managed to get is still significantly short of even one MByte per second I'd have to say I'm skeptical that bus performance is really a blocker here, given even plain old Nubus should easily handle at least ten times that even using the slowest transaction modes.

 

[beachycove]

The closest Apple got to a machine that meets that description is the faster versions (100mhz) of the 8100...

Would the AWS9150 be the same? I have the last iteration here, running at 120mhz.

 

[Gorgonops]

Would the AWS9150 be the same? I have the last iteration here, running at 120mhz.

Wow, I was actually able to find the devnote for the WS9150 floating around out there, and according to it the formula is the same as the 8100; machines slower than 100mhz have the "BART4", newer machines have the "BART21".

The one difference between the two that looks like it might be relevant for testing an Ethernet card is the BART4 has a limitation with burst mode; it's all or nothing, if you have any Nubus cards plugged in that don't support it then you can't enable that transfer mode for the one that does. BART21 can enable burst mode per slot.

 

[beachycove]

Woohoo! I have a new reason to live fire up that 9150! And in that same spirit, say I had a Radius LeMans, or a Jackhammer, or something equally exotic. Could I expect them to perform markedly better in their own right in the 9150 than in, say, a Quadra 650? It’d be interesting to try and measure the difference somehow, as has been happening in this excellent thread.

 

[jeremywork]

Aha! Found an explanation for the RAM disk problem I was encountering:

• The Macintosh Quadra 950 cannot deal with RAM disks if 256 MB (the upper limit) of RAM is installed.
  

source: https://apple.fandom.com/wiki/RAM_disk

Interestingly, this is not an issue when the PPC is running. I'll be able to pull out a bank and give the 68040 a fair run now.

However, I was able to use RAMometer 1.3.4 to calculate the actual access rates on both platforms, as well as the IIfx for reference.

Quadra 950 (68040):

6,405 KB/s

Quadra 950 (PPC):

15,668 KB/s

IIfx (8x4MB):

2,912 KB/s

This said, at this point I'm not expecting the 68k test to outpace the previous PPC one, and I'm now confident the "16.7 MHz System Bus" reported by Metronome is erroneous.

In the meantime, the AAUI to 10Base has been consistently faster than the Asanté ever since the sw update. When I have my QS02 up and running, I'll be excited to see if AppleShare IP server brings any good news...

I wish my 840av would awaken from its tomb... I could test it in the IIfx (not expecting much but at least one slot is used up for ethernet regardless) but I expect this was probably intended for use on a BART21 system, as Gorgonops suggested. 

I'd love a 9150, but an 8100/100 is probably a more attainable test subject  ;-)

 

[Gorgonops]

However, I was able to use RAMometer 1.3.4 to calculate the actual access rates on both platforms, as well as the IIfx for reference.

Quadra 950 (68040):

6,405 KB/s

Woah. I wonder exactly what this RAMometer program is doing as the basis of its bandwidth calculations. As I've grumbled recently in another thread I think people sometimes forget just how slow these machines are by modern standards, but 6.5MB/s is pretty slow even by my jaded estimates of how fast a Quadra's memory subsystem might be. (The "almost triple that" the PPC version gets hitting the same memory suggests to me that there's *something* CPU expensive about it pushing the number down.)

 

[Cory5412]

I actually have a 9150 hanging around at the moment, it's for @wthww. I also have a working 840av (though I was hoping to economize on potential teardowns/rebuilds to preserve its plastic) as well as a 950 (also for wthww) and a 6100.

I would be surprised if machines for lending or someone willing to meet up to test the card wasn't a bit closer though.

 

[jeremywork]

Woah. I wonder exactly what this RAMometer program is doing as the basis of its bandwidth calculations. As I've grumbled recently in another thread I think people sometimes forget just how slow these machines are by modern standards, but 6.5MB/s is pretty slow even by my jaded estimates of how fast a Quadra's memory subsystem might be. (The "almost triple that" the PPC version gets hitting the same memory suggests to me that there's *something* CPU expensive about it pushing the number down.)

I'm not sure exactly how it's calculated (obviously the cleanest test is to transfer data from something faster, like the PPC's 1MB L2 cache or I suppose the 68040's 4KB data cache.) RAMometer seems to write a pattern to memory, then read the whole thing back and test the size of the written file over the time it took to complete the I/O.

I have a second 950 that I'd like to test with to confirm there was nothing wonky going on in this test (the 950 with PPC has some oddball issues, such freezing instead of powering down or restarting after finder unloads, but only when booted from the 68040, and even with the PPC and Nubus physically removed.)

I actually have a 9150 hanging around at the moment, it's for @wthww. I also have a working 840av (though I was hoping to economize on potential teardowns/rebuilds to preserve its plastic) as well as a 950 (also for wthww) and a 6100.

I would be surprised if machines for lending or someone willing to meet up to test the card wasn't a bit closer though.

If you come across any leads for another 9150, I'd definitely be looking to purchase.

I'm not happy about spending as much as I did on a recapped 840av board only to have it start up exactly twice, at which point I left my positive feedback. I still have the original board too, but it produces a death tone after the chime no matter what I do... If only the 950 had that nicer bus chip, they're extremely reliable in my experience.

If I could find a 9150, I'd let the 840av project go to someone with a higher pain tolerance. 

Now that I think of it, I never did try a different PSU... maybe that's my issue?

 

[trag]

Woah. I wonder exactly what this RAMometer program is doing as the basis of its bandwidth calculations. As I've grumbled recently in another thread I think people sometimes forget just how slow these machines are by modern standards, but 6.5MB/s is pretty slow even by my jaded estimates of how fast a Quadra's memory subsystem might be. (The "almost triple that" the PPC version gets hitting the same memory suggests to me that there's *something* CPU expensive about it pushing the number down.)

I don't think that number is meant to be raw RAM bandwidth.  I think it is RAMometer's report of the speed it managed in testing.   A substantial portion of the test process is all the compare and branch statements.  Which would explain why the PPC is so much faster.

A typical test sequence involves writing the test pattern to memory.   Then reading a portion to RAM, or just reading one word to a register.   Then compare what was read to the test pattern, branch to the next test on pass, or fall through to error on a fail.   Add in increments to at least the read address counter.

So, at the least you have an address increment, a memory write, another address increment, a memory read, and a compare per every single word of memory tested.   That's a far cry from how long does it take to read one word; how long does it take to write one word.

BTW, I got similar results back when I was testing IIfx RAM with RAMometer:

In theory

 

[jeremywork]

A typical test sequence involves writing the test pattern to memory.   Then reading a portion to RAM, or just reading one word to a register.   Then compare what was read to the test pattern, branch to the next test on pass, or fall through to error on a fail.   Add in increments to at least the read address counter.

Reading that number as "test performance" makes sense. Knowing how much a 68040 benefits from having cache installed is a clear indication that it completely saturates the bandwidth of the memory, there should be no effective difference between the access speeds of the two processors. The main reason I ran the test was to see if the PPC had slowed the bus down in order to time it's 100MHz clock, so I wanted to make sure the number wasn't lower.

Do you know what the best way of testing RAM bandwidth is on 68k? Gauge Pro has a 'moving memory' test but only runs on PPC.

 

[trag]

I'm not aware of a RAM bandwidth tester.   That doesn't mean I've never seen one.  It just means I can't remember seeing one.   Today.   If I was looking, I might take a look in Tech Tools Pro.    It seems like that's the kind of thing that might be in there, but I don't remember ever seeing one.

 

[jeremywork]

AsantéFAST 10/100 Nubus Update:

I was able to test this card for a while with a finally-running Power Mac 8100/100 [one of the machines to use the final-revision BART21 Nubus controller.] Unfortunately the ATTO SE IV I installed worked at first, but became less and less stable and now causes the machine to lock at bootup (death tone on the Q950; swapping for the one I usually run in the 950 resolves the issue) Despite that, I was able to reproduce the same benchmarks I ran on the 950 in previous posts.

Running Mac OS 8.6 on the 8100, the disk performance to the 68-pin drive on the SE IV was about 14-15MB/s in sustained writes (ATTO performance utility)

Installation of the AsantéFAST driver posed no issue, and once again I used the 9600/1000 running 8.6 with its 10/100 PCI card as the appletalk share. 

As I found previously on the Q950, despite enabling RAM disks on both machines, the fastest speeds were always achieved by transferring directly from SCSI drive to SCSI drive, facilitated by the destination machine, not the source.

The fastest speed I could consistently achieve on the 100MHz 601+ was about 760KB/s when copying from my 9600's UW drive (mounted on the 8100, copying on the 8100)

I then installed a Sonnet Crescendo 266/1M in the 8100. Despite information I found online, the Sonnet upgrade had no problem booting from the ATTO SE IV. Asanté drivers continued to work as well.

In the same test scenario, the fastest I could consistently achieve on the 266MHz G3 was about 900KB/s.

I finally installed the fastest possible CPU in the 8100, the Sonnet Fortissimo 500/1M. Again, no conflicts in this configuration (though I should mention the HPV ribbon cable apparatus sometimes needs to be reinserted a couple times at all joints, or you may get distorted or no video from the HPV card.)

The same test on the 500MHz G3 netted a consistent transfer rate of 1.28-1.31MB/s. Enough to break the 10Base/T barrier, if only by a technical margin (though considering overhead, the gain is slightly more significant.)

I still hope to revive my 840AV soon to see if the faster memory system does anything here, but given what I've found so far, this card is looking less and less useful. I'm also still working on getting my Quicksilver properly running OS9 to get ASIP running (7447 upgrade doesn't seem to like some extensions in the OS9 base pack, aside from the documented multiprocessing folder; haven't fully investigated yet.) Once those tests can both be run I'll be happier to call these results definitive, but figured I'd update this in the meantime as I don't have an explicit ETA on either of those projects.