shamino wrote on 2025-01-30, 07:00:

On some systems the PCI bus gets saturated. Also, the frontside bus and RAM can get hit 4x: reading a file from disk to RAM, reading that file data back from RAM, converting it to network packets as it's written back to another location in RAM, and then reading those packets back out of RAM to send them out to the network.

I just replaced the 3Com network card in my Celeron 600 rig with an Intel PRO/1000 GT. Using that card, and with "Interrupt Moderation Rate" disabled in the driver, I'm getting the following results from LAN Speed Test under Win2K:

However, in real-world scenarios such as copying an ultra compressed 7-Zip archive from the network, I'm getting around 13.1 MB/s. Uploading to the network is faster at around 17.6 MB/s. So I think I've hit the practical limits of what is possible on this system. It also seems consistent with the results from this thread: There is a difference between Intel Pro 1000 NICs

By the way, drivers for the Intel PRO/1000 GT are quite resource heavy on this Celeron 600 rig. And my motherboard always loads the network card's boot rom at startup, only then proceeding to boot from the local hard drive. Because of that, I might go back to the lightweight 3Com card.

Here is my experience, which I am quoting from another thread

EDIT: The picture attachments are in the wrong order for some reason. I will fix it ehen I have access to a PC. In the meantime, they are correct in the original thread/post

darry wrote on 2021-01-17, 01:49:

I was feeling bored today, so I dug up my 400MHz Pentium 2 (4x100MHz), installed it in my Asus P3B-F and ran at 266MHz (4x66MHz) […]
Show full quote

I was feeling bored today, so I dug up my 400MHz Pentium 2 (4x100MHz), installed it in my Asus P3B-F and ran at 266MHz (4x66MHz) to get something fairly comparable to your CPU and then ran it with a 1400MHz CPU for a best case scenario .
- The machine was running Windows 98 SE and was summarily TCP optimized using https://www.speedguide.net/files/sguide_tweak_9x.zip (TCP Optimizer set to optimum did not give optimal results) .
- CPU-Z was run after the FTP transfer had completed

First, a disk benchmark with the 266MHz CPU :

The attachment 266mhz_atto.png is no longer available

This is running in DMA mode, obviously, off of a SIL3114 SATA controller

Second, the best performance I could get with a 266MHz CPU through FTP using a 3COM 3C905B (100Mb NIC):

The attachment 266mhz.png is no longer available

Third, the best performance I could get with a 266MHz CPU through FTP using an Intel PRO/1000 GT (Gigabit NIC) :

The attachment 266mhz_gig.png is no longer available

Finally, the best performance I could get with a 1400MHz Tualatin (PowerLeap adapter) CPU through FTP using an Intel PRO/1000 GT (Gigabit NIC) :

The attachment 1400mhz_gig.png is no longer available

It should be obvious that a 266MHz Pentium II CPU is just not fast enough to drive a Gigabit NIC .

Conclusion :
Assuming my results are representative, running a Gigabit NIC rather than a 100Mbps one on a 266MHz Pentium II CPU is practically pointless.

EDIT : corrected a typo and an image
EDIT2 : Added some context

darry wrote on 2025-01-30, 13:57:

Assuming my results are representative, running a Gigabit NIC rather than a 100Mbps one on a 266MHz Pentium II CPU is practically pointless.

Interesting findings.

I think Slot 1 (and Socket 370) systems are probably where one would want fast network transfer speeds the most (even if it's "just" 100Mbps), because they only have USB 1.1 by default. And adding a USB 2.0 card has the potential to slow down the entire system, depending on a number of factors.

over 7 mb/s networking using a p2 266 on 9x isn't too bad in my book. that's actually around the sequential speed of some decent IDE drives just a year older than that CPU. for more serious work NT would have been used and would perform better than that. i think an interesting use case for fast networking is to stream a disc image using iSCSI, but this needs win2k anyway.

the pro/1000gt and probably also the 3c905b support checksum offloading, but i forgot if the drivers under 9x expose this. if they don't, this is probably not supported under 9x at all.

Joseph_Joestar wrote on 2025-01-30, 14:10:

darry wrote on 2025-01-30, 13:57:

Assuming my results are representative, running a Gigabit NIC rather than a 100Mbps one on a 266MHz Pentium II CPU is practically pointless.

Interesting findings.

I think Slot 1 (and Socket 370) systems are probably where one would want fast network transfer speeds the most (even if it's "just" 100Mbps), because they only have USB 1.1 by default. And adding a USB 2.0 card has the potential to slow down the entire system, depending on a number of factors.

To be fair, that system does reach something like 400Mbps under Linux AFAICR. I do not see a point to running XP on that slow a system for my use cases.

Other than the challenges with USB 2.0 and system slowdowns on older platforms, I don't like the idea of plugging/unplugging stuff constantly. I'm lazy that way.

auron wrote on 2025-01-30, 14:31:

over 7 mb/s networking using a p2 266 on 9x isn't too bad in my book. that's actually around the sequential speed of some decent IDE drives just a year older than that CPU. for more serious work NT would have been used and would perform better than that. i think an interesting use case for fast networking is to stream a disc image using iSCSI, but this needs win2k anyway.

the pro/1000gt and probably also the 3c905b support checksum offloading, but i forgot if the drivers under 9x expose this. if they don't, this is probably not supported under 9x at all.

Yeah, back in the day, with slower drives, the bad network performance of Windows 9x was not that noticeable for home users.

I noticed that the Intel PRO/1000 Network Adapter Driver 9.2 WHQL is still hosted here on Softpedia.

That driver version was released in either 2004 or 2005. It certainly feels much lighter on my Celeron 600 under Win2K than the 14.0 drivers that I was using before. The installer is a bit wonky, but it does work if you update the driver manually via Device Manager. Transfer speeds remain mostly the same as before, so it might be a good fit for retro systems with less horsepower like my Celeron rig.

Joseph_Joestar wrote on 2025-01-30, 13:30:

And my motherboard always loads the network card's boot rom at startup, only then proceeding to boot from the local hard drive.

I found a way to disable this. It can be done using Intel's BOOTUTIL.EXE and I specifically needed the DOS version for my retro rig. I got it from a driver CD image containing Intel Network Connections Software Version 16.2, but it can probably be found elsewhere as well. The syntax is as follows:

1BOOTUTIL.EXE -FD -ALL

This will disable the boot rom on all Intel network cards in the system. Worked for me, so no more Intel Boot Agent during startup on my Celeron rig.

I did some more network testing today, by adding a 3Com 3C905B-TX network card to my K6-2 (non-plus) 266 build.

Under Windows 95 OSR 2.1 I was getting fluctuating file transfer speeds between 2-4 MB/s. However, under Windows NT 4.0 + SP6 I was getting a steady file transfer speed of 5.5 MB/s. Once again, the NT network stack proves its superiority. Also, both Rx and Tx offloading is available for this card under NT4 when using the latest official drivers.

I tested with iperf. The client was the Win98 box in my signature, with a 3Com 3C905C-TXM NIC. The server was a Ubuntu 24.04 LTS virtual machine with my NAS as a host. The Win98 box connects via a Unifi Device Bridge to a U6 Mesh AP that is hardwired to the LAN at 1 GBe. The NAS is hardwired via 10 GBe. TCPOptimizer set to Optimal did not change the results.

I have many benchmarks of Samba share speed on old CPUs. You can find them here:

ctrl+f: Total Commander 8.51a – Samba network drive

http://hw-museum.cz/article/5/cpu-history-tou … -1995---1999-/9
http://hw-museum.cz/article/6/cpu-history-tou … -1999---2001-/9
http://hw-museum.cz/article/8/cpu-history-tou … -2001---2003-/8

havli wrote on 2025-02-13, 15:08:

I have many benchmarks of Samba share speed on old CPUs. You can find them here:

Those are some nice results.

From what I understand, the server was Debian Linux. But what about the client OS from which you used Total Commander? Was it Win98SE or something else?

It was Windows XP SP3 in all these tests.
And Marvell Yukon 88E8003 PCI GLAN for all platforms.