You understand that you are not measuring the ISA bus speed here, but the overall performance of the system, right? Is that a fast 486 with on-chip L1 talking to the bus, or something else? Pure DOS/BIOS with it's code path meant to run on 8086, or some optimized device driver?

Deunan wrote on 2024-07-25, 16:12:

You understand that you are not measuring the ISA bus speed here, but the overall performance of the system, right? Is that a fast 486 with on-chip L1 talking to the bus, or something else? Pure DOS/BIOS with it's code path meant to run on 8086, or some optimized device driver?

So I'm working with an AMD 486 @ 133 Mhz w/ 0 wait state memory, 256KB Cache, shadowed ROMS, no Himem or EMM386.

All ISA hard drive transfers at <= 2MB/s when using a non caching ISA storage controller, even when reading from the drive cache. This is with the stock BIOS, MrBios, XTIDE, SIIG Enhanced BIOS, EZDrive overlay, etc.

When using a caching ISA IDE controller, it can read at 4MB/s from the controller cache, but gets something like 1.4MB/s when reading directly from the drive, as is the nature with those caching cards.

The computer is fast enough to transfer at 4MB/s. The ISA is fast enough to transfer at 4MB/s. The drives are fast enough when operating in a fancier ATA mode.

Just seems like it must be ATA overhead that is keeping things a lot lower than they theoretically could be, but maybe I don't understand the issue.

ISA IDE = PIO mode 0 = 3.3 MB/s
A drive, however, can't push the data all the time - often it's waiting for the heads to move instead.
Overall, it's much less than 3.3 MB/s.

A caching controller may also use PIO mode 0 at the ISA side - again, 3.3 MB/s.
But it may also use bus mastering - possibly faster.
At the HDD side, such a controller may also use something faster than PIO mode 0.
And when the drive is busy moving heads, the controller can still push data from the cache.

Also, a mechanical drive (depending on it's age) might just not be able to read the data from the platters any faster than a few MB/s. The older MFM drives will only go up to 350 kB/s or so.

The reason I asked what code is running is already explained, DOS/BIOS will use PIO 0, but even at higher PIO modes and faster CPU you are still limited by the I/O performance. My guess is that caching IDE controller lets you access the data as a memory mapped to ISA space? It would not be able to push 4 MB/s otherwise, not through I/O.

Are you sure the caching IDE still uses programmed I/O to transfer data between the system and cache? Or could a shared memory region be involved?

Deunan wrote on 2024-07-25, 17:16:

My guess is that caching IDE controller lets you access the data as a memory mapped to ISA space? It would not be able to push 4 MB/s otherwise, not through I/O.

jakethompson1 wrote on 2024-07-25, 18:31:

Are you sure the caching IDE still uses programmed I/O to transfer data between the system and cache? Or could a shared memory region be involved?

That makes sense. Takes fewer cycles to do MMIO than it does to do PIO. It comparing PIO to MMIO seems similar to comparing Sata Legacy ATA with Sata AHCI (or NVME even)

Edit: Doing the math, MMIO comes out to about 4.2MB/s max throughput while PIO comes out to about 3.3MB/s max throughput. I can believe that the ATA & IRQ overhead reduces the throughput from 3.3MB/s to 2MB/s much more easily than going from 4.2MB/s to 2.0MB/s.

Deunan wrote on 2024-07-25, 17:16:

Also, a mechanical drive (depending on it's age) might just not be able to read the data from the platters any faster than a few MB/s. The older MFM drives will only go up to 350 kB/s or so.

Yeah I've known "classic" 486 era IDE drives to top out at 1500kB/s then later 1 and 2GB but early for that size, only doing 2500.