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

[Trash80toHP_Mini]

If you're getting anything at all over 12MBps across SCSI you're kicking the teeth out of a 100Mbps Ethernet card. That's not including handshaking inefficiencies of Ethernet. SCSI transfers are just plain faster. At 3MBps+- the crappy mobo SCSI on early Macs you're still talking 24Mbps, no?

 

[jeremywork]

If you're getting anything at all over 12MBps across SCSI you're kicking the teeth out of a 100Mbps Ethernet card. That's not including handshaking inefficiencies of Ethernet. SCSI transfers are just plain faster. At 3MBps+- the crappy mobo SCSI on early Macs you're still talking 24Mbps, no?

Ah, to clarify, I get 14-15MB/s when benchmarking the disk with ATTO performance utility, which is simply a CPU-Disk test, no network involved. Copying files within the drive is very fast, but network transfers on the 100Mbps card never exceeded 1.3MB/s (and that required a 500MHz G3 to achieve.)

 

[Trash80toHP_Mini]

1.3MBps would be 10.4Mbps, that's really disappointing all things considered, but still probably a lot better than vanilla 10bT on the same systems, no?

 

[jeremywork]

10.4 is pretty underwhelming, but seeing as throughput seems to scale with CPU speed I'm glad to know it CAN exceed 10Mb/s. Tests on the 950 showed that especially in later AppleShare client packages, the onboard AAUI was faster than Nubus cards, so I'd expect the 8100's AAUI to get near its theoretical maximum in this case. I really want to figure out what's wrong with this SiliconExpress so I can get back to testing on it, but for now it's back on the shelf awaiting some time with a multimeter.

 

[Trash80toHP_Mini]

I'd expect the 8100's AAUI to get near its theoretical maximum in this case.

That's still slower than 10.4 and handshaking takes a big bite out transfers so you're probably a lot better off with the 10/100 NuBus Card. Even Beige G3 10bT will be noticeably slower than that, no? Forget SCSI transfer rates and test Ethernet to Ethernet implementation transfer ratesw and you'll probably feel a lot better about that 10/100 NuBus Card. []

 

[jeremywork]

There's certainly an improvement, but I'll be curious to get exact numbers. What constitutes an Ethernet to Ethernet transfer test? In my experience, large single-file disk transfers over Appletalk are the fastest I can get the network to measurably move.

This speed roughly matches the one test I've done with a couple G3 Gazelle systems with Farallon CSII 10/100 cards. I only had my 2012 MBP to facilitate a connection to a wifi network in this case, but with two NICs, each serving one system via Internet Sharing and each with a 100mbps link. If I use the 500MHz system (a TAM) to download a file over HTTP, I can saturate about 4Mbps, while I use appletalk to transfer another file from the 400MHz system (a 5500) I see bursts of up to 11Mbps on the TAM. I intend to re-run these tests on a standard wired network, but it has me thinking these earlier machines just don't network very fast.

(5500/top; WAN; TAM/bottom)



The 9600 in its current configuration can serve data at 2.9MB/s to a Mystic G4/1.5GHz on 9.2.2; transferring back from this machine to the 9600 yields 3.6MB/s. I would expect the two Gazelle machines to be closer to this range with their G3/Farallon configuration, so perhaps my laptop was responsible for overhead on that test, or perhaps, like the 8100 @ 500MHz, that's just as fast as it can go.

 

[Trash80toHP_Mini]

I don't bother with benchmarking. Early machines run period specific software as fast or faster than later machines running their period specific software. Featuritis and bad code outpace Moore's law, 'nuf said.

 

[jeremywork]

I agree that the speed of operation has more to do with software efficiency than raw throughput, but that doesn't mean I'm not curious to see the numbers  

In this particular case I was more curious to see whether this card is worth taking the slot it uses. In the 950, even with a PowerPro, the answer was 'save the slot, the AAUI is just as good/better' In the 8100/500 I'm happier with it, though certainly not as dramatic an advantage as I would've imagined.

 

[Trash80toHP_Mini]

Faster networking in a slot deprived Mac (x100 series) even at a full throttle 100Mbps would be a terrible waste of a precious slot for any but niche applications. That said, something strikes me as fishy about your results, else why would 10/100 cards have been made for NuBus Macs?

 

[jeremywork]

something strikes me as fishy about your results, else why would 10/100 cards have been made for NuBus Macs?

This is what I keep thinking; I was especially underwhelmed by the speed on the stock 100MHz 601, only about 20% faster than the Q950 with the same processor. Nubus 90 is definitely working, as I get double the disk I/O from the same ATTO SE IV tested in the Q950.

Gauge Pro's 'Moving memory' test also showed nearly 3x the speed of the Q950:

Quadra 950 @ 100MHz PPC - 12MB/s

Power Mac 8100/100 - 30MB/s

Power Mac 8100/266 - 34MB/s

Power Mac 8100/500 - 38MB/s

The fact that it does get faster with faster machines makes me wonder if something in the driver is just horribly inefficient...

 

[jeremywork]

I'm not sure how I overlooked this for so long, but the Cnet product page reveals in the title this was shipped with something called Netdoubler:

https://www.cnet.com/products/asantefast-10-100-nubus-w-netdoubler/

And here it appears to be. Looks very promising.

https://macintoshgarden.org/apps/netdoubler

Now I regret reassembling the 950 with its fragile case door... hopefully I can get that broken ATTO working so I can test this on the 8100 soon (this is partially because I ran out of 50 pin drives and I don't have an adapter handy. I'll probably get one of these and fall back to the 10MB SCSI on the 8100.)

 

[Trash80toHP_Mini]

Most excellent news. What's the copyright date on the card? That will tell you what would be contemporaneous production machines.

 

[Cory5412]

That said, something strikes me as fishy about your results, else why would 10/100 cards have been made for NuBus Macs?

My guess is entirely that people thought it was important and it was "cheap" to build, so someone did, even though there's extremely no reason to get it.

That you need to drop a G3 CPU in to even exceed what onboard 10 megabit can do is indicative, of how impractical this upgrade is in the real world. It's a checkbox item, not a real performance boost in any meaningful way.

I don't bother with benchmarking. Early machines run period specific software as fast or faster than later machines running their period specific software. Featuritis and bad code outpace Moore's law, 'nuf said.

Sure, but, however, comma: This is a scenario where using a benchmark can help set expectations on the performance of a specific subsystem.

In-situ, benchmarks are used to determine the viability or worth of a particular component upgrade or, for example, how much computer to buy at any given moment.

The gotcha, of course, is that it's important to know what the critical subsystem for your application is. Buying a RAM upgrade doesn't make sense for an app whose critical path is disk and buying a network upgrade doesn't make sense for an application whose critical path is CPU, for example.

To that end:

That's still slower than 10.4 and handshaking takes a big bite out transfers so you're probably a lot better off with the 10/100 NuBus Card. Even Beige G3 10bT will be noticeably slower than that, no?

10.4 vs 10 megabits isn't an awful lot.

Either some particular software had to have had some really special sauce going in or these cards just wouldn't have been worth it except to check a box, even though everyone knew the box being checked wouldn't at all be meaningful.

Somewhere up -thread, I talked about an FDDI add-in and the "big speed uplift" it provided for big file transfers using a special file transfer utility meant to work with that card's network protocol, and the reaction it got was essentially "that's still a miserable file transfer rate" - and, I mean, even 13 megabits per second would still be.

It mostly comes down to a combination of Classic Mac OS being exceptionally bad and the hardware being pretty ho-hum.

I've seen some reports that even on plain Ethernet, there were different networking tools (Novell IPX for example) that are alleged to have been faster, but I haven't/don't have the resources to test that.

 

[jeremywork]

My guess is entirely that people thought it was important and it was "cheap" to build, so someone did, even though there's extremely no reason to get it.

That you need to drop a G3 CPU in to even exceed what onboard 10 megabit can do is indicative, of how impractical this upgrade is in the real world. It's a checkbox item, not a real performance boost in any meaningful way.

I had always considered this a possibility; after seeing 10/100 talked up (here and other places too) as a viable upgrade for anything with a Nubus slot it was the only category of common 'accelerator' I couldn't find published results on. I'm lucky to have found this card for $30, so I wanted to see whether "checking that box" does anything other than consume a Nubus slot (and Apple's A/ROSE card may still be a better choice for most Nubus machines, despite being only 10Mbit.) Needing to use the fastest G3 available in order to see such a tiny benefit was not what I expected, but gives better context to my first results with the Q950. Hopefully there turns out to be some secret sauce in NetDoubler, or some other scenario I haven't considered, or this will turn out to have been colossally overhyped

I'm not even sure there are compatible drivers, but perhaps this was better in A/UX environments?

What's the copyright date on the card?

©1995 Asante Tech, Inc

By this estimation, it may have *only* been intended for use in PowerPCs, though their driver runs fine on a 68040 as well in my experience.

 

[Cory5412]

I misread something along  the way and failed to realize that 10.4 megabits (1.3 megabytes) per second was the top you achieved with the G3/500 (if I'm not misreading again) which is definitely that much less ideal.

There's gotta be some kind of secret sauce in that netdoubler software, or some kind of, server software this could've been used with on an AWGS8150/9150 where 100 megabits would have made $PRICE_OF_THE_CARD of a difference.

A 1995 copyright date is a fair bit before 100mbit was starting to become "cheap" (to be honest it still wasn't "cheap" in 1998) and to make matters worse, network switches that we rely on for modern high performance ethernet network were nearly unheard of in 1995, and using this thing on a busy hub would have made observable performance that much worse than it is on your switched network at home today.

One possible interesting thing would be to see whether transfers between system 7/8/9 and, say, a netware or windows 2003 server might be faster. I don't have any netware stuff, but NT tended to do better with services for macs and so you might be able to show some better numbers that way.

 

[jeremywork]

10.4 megabits (1.3 megabytes) per second was the top you achieved with the G3/500 (if I'm not misreading again) which is definitely that much less ideal.

Yep, you read that correctly. 

There's gotta be some kind of secret sauce in that netdoubler software, or some kind of, server software this could've been used with on an AWGS8150/9150 where 100 megabits would have made $PRICE_OF_THE_CARD of a difference.

This is what I kept coming back to... This wasn't a cheap compatibility upgrade from the early 2000s; they had to sell people on this in 1995. Your point about real-world testing against a heavily loaded shared hub makes sense.

One possible interesting thing would be to see whether transfers between system 7/8/9 and, say, a netware or windows 2003 server might be faster. I don't have any netware stuff, but NT tended to do better with services for macs and so you might be able to show some better numbers that way.

Not really set up for this at the moment, but next time I have the thing out and running I'll set up a windows server and see.

 

[Unknown_K]

I have some 10/100 Mb Nubus cards and while the speed is better then 10Mb the classic Mac OS really slows things down plus I don't know if there is DMA transfers involved or if there is cache on the Ethernet card.

If you want to see a home network fly try running A/UX on a Q950 with the PDS cache card installed, probably 2x or more faster then classic OS on the same built in Ethernet.

 

[trag]

IIRC, Net Doubler only works between two machines which are each running it (and have installed compatible Asante hardware?).   I always assumed it was some kind of compression before transmission software, but that's probably wrong.   Hardware of the time probably wasn't able to compress on the fly that well.

 

[jeremywork]

I have some 10/100 Mb Nubus cards and while the speed is better then 10Mb the classic Mac OS really slows things down plus I don't know if there is DMA transfers involved or if there is cache on the Ethernet card.

If you want to see a home network fly try running A/UX on a Q950 with the PDS cache card installed, probably 2x or more faster then classic OS on the same built in Ethernet.

I'm curious what improvements you can discern with yours, and in what machines the difference can be felt. I do intend to test in A/UX if I can find suitable drivers.

IIRC, Net Doubler only works between two machines which are each running it (and have installed compatible Asante hardware?).   I always assumed it was some kind of compression before transmission software, but that's probably wrong.   Hardware of the time probably wasn't able to compress on the fly that well.

Interesting, hopefully I can find another machine that it'll run with. I probably have some Asante PCI cards around... If it does rely on compression, I suspect only PPCs would theoretically be able to keep up in a way that would net a meaningfully faster transfer. 

 

[jeremywork]

Looks like there is some cache; not sure if there's anything to offload compression/decompression.