Here my test result with a Cyrix 5X86@120 /w 3COM 3C509-B

DOS with mFTP and things like netware client or EtherDFS : ~700KiB/s
Windows 95 & NT4 and Linux : ~300-400KiB/s MAX, usually less than that.

I suspect Full duplex on Windows/Linux does not work with that network card for some reason.
I can't find a way to turn it on on Windows, and when I turn full duplex on on Linux, most of the packets get lost.

In pure DOS, using mTCP and spdtest, DE22XPD and 3C90XPD drivers.

Having the P233 be the sender and the P120 be the listener, it reaches 8 Mbps (1000 KB/s),

Switching places, P120 sender, P233 listener, I get 5.3 Mbps (662.5 KB/s).

P120 mTCP FTP download yields 6.82 Mbps (853KB/s), I guess there's losses due to overhead for FTP or it could be my FTP server expects faster clients as pointed out before.

But for this machine, I think that's as close as I'm going to get.

Update:

With the Cyrix 486-66 using mTCP FTP and the Realtek 8019 Packet Driver (which looks identical to the Crynwr driver btw) I reached 6.4 Mbps (800 KB/s)

I think this is as fast as they're both going to get, though I am still surprised that the P120 still couldn't max out the card, but maybe it's not the CPUs fault. Who knows?

Thanks to everyone who replied, I'm still interested to see what everyone else "benchmarks" on low end machines and 10 Mbit ISA cards though.

Maybe we can build a matrix of card, driver and CPU with speeds to record our findings?

Twisted.

Rav wrote on 2024-01-28, 17:49:

Here my test result with a Cyrix 5X86@120 /w 3COM 3C509-B […]
Show full quote

Here my test result with a Cyrix 5X86@120 /w 3COM 3C509-B

DOS with mFTP and things like netware client or EtherDFS : ~700KiB/s
Windows 95 & NT4 and Linux : ~300-400KiB/s MAX, usually less than that.

I suspect Full duplex on Windows/Linux does not work with that network card for some reason.
I can't find a way to turn it on on Windows, and when I turn full duplex on on Linux, most of the packets get lost.

Thanks very much for your tests, gives me a clue as to what expect realistically.

TwistedSoul21967 wrote on 2024-01-28, 17:48:

However, going back to the unmanaged yielded the previous result of 3.9Mbps (490KB/s) which is very odd considering auto detect is off so it should still use 10/FD.

Here's your problem. The unmanaged switch will link at 10/HD by default, and you are forcing FD at the NIC.

So I did some tests...

CPU: Celeron 266
NIC: Novell NE2000plus3, using the RJ45 port, no Full Duplex support

all tests using FTP client:
bin
get testfile.bin nul

Windows 98SE:
"NE2000 Compatible" driver
1000..1005 KB/s

mTCP:
NE2000 11.4.3 packet driver
MTU 1500
907..917 KB/s

One thing to keep in mind is that the FTP client writing to NUL still has extra overhead because it's probably doing the byte copies to the NUL device.

If you want to know the upper limit for an adapter run a benchmark that just sends bytes but doesn't simulate writing them. I wrote the spdtest program for mTCP to do that because even on a slow machine, those extra copies are measurable overhead. (Spdtest does everything a socket connection is supposed to, including verifying the checksums, handling retries, etc. But it just throws the data away instead of copying it.)

I think between my previously collected data points on the mTCP performance page and the new data points here we've conclusively proved that an ISA Ethernet card in a reasonable system can saturate a 10Mb/sec Ethernet connection.

I was able to achieve 980 KB/s though http download (lan) using pentium 166 mmx on FIC PA 2002 and 64MB EDO ram under windows 95 C.
The nic was an ISA version of 3com etherlink (very reliable card working stable).

TwistedSoul21967 wrote on 2024-01-28, 18:16:

I think this is as fast as they're both going to get, though I am still surprised that the P120 still couldn't max out the card, but maybe it's not the CPUs fault. Who knows?

Twisted.

You are never going to max out the card. The line speed of the card does not take into account protocol overhead, collisions/broadcast on the segment, ISA bus timing (clock speed, wait states, congestion) etc etc the list is a mile long.

8 Megabits FTP throughput is a perfectly reasonable "max" for a 10 Megabit ISA card.

So I wanted to know what's the minimum CPU to saturate the 10 Mbps Ethernet...

CPU: 386DX-40
NIC: 3C515 (100 Mbps)
software: mTCP with MTU 1500

First, ran 3C515CFG.EXE, did Auto Configure, including the "tune the DMA parameters" procedure.
FTP downloading to NUL:

956..957 KB/s

Rebooted, and switched it to de-turbo:

CPU: 386DX-20

Again, 3C515CFG.EXE, Auto Configure, tune the DMA parameters.
FTP downloading to NUL:

552..572 KB/s

Conclusion:
- low-end 386 can't use 10 Mbps to the max
- high-end 386 is already fast enough to saturate 10 Mbps
- it probably doesn't make sense to use a 100 Mbps NIC in this machine, the results aren't really better than 10 Mbps