IDE Co-Processors?

morbid asks: "EIDE is generally considered to be inferior to SCSI because it requires more involvement from the processor slowing the system down, but would it not be possible to build an EIDE/ATA (?) controller with its own processor, freeing up the CPU and increasing system performamce while allowing the use of inexpensive drives?"

Comments

Say hello to Bus Mastering

[gammatron]

Most modern motherboards (probably all) support bus mastering, which does what you want.

Bus Master IDE FAQ

Bus Mastering IDE technology implements logic circuitry on your motherboard to reduce CPU's work of retrieving and storing data on your hard disk drive. This technology can potentially "free up the CPU" to do other tasks in a multitasking operating system environment such as Windows* 95, Windows NT*, OS/2* etc.

--

Re:Say hello to Bus Mastering

[Tower]

Yes, bus mastering has been around since the PIIX days (95/96), but all that says is that the IDE controller is a bus master on the PCI bus, with DMA capability. Whoop-dee-freakin-doo. This still doesn't let the main proc out of the other tasks of drive management that it has to do. There's a lot of good stuff that a SCSI initiator and target take care of that the IDE controller and target don't, and this is where the benefit is.

So, yes, bus mastering is good. My $3 network card does it too, as does (should) any non-brain-dead PCI device that involves any sort of data flow (LAN, drive controller, video, capture). It doesn't take care of everything, though.
--

Re:Say hello to Bus Mastering

[Paulo]

Well, I have. Using Windows NT (cough) with normal drivers, copying a large file (40-50Mb or more) would take 100% of processor time, while with bus mastering drivers it took just 3%-5%, or even less.

SCSI & Multi-tasking OS's

[maggard]

Well, SCSI is a bit more then a fast way of talking to a drive. It also frees the CPU from many of the drive geometry issues that it would otherwise have to track. SCSI also allows things like multiple devices per channel, request queueing, smarter IO, disconnect/reconnect, etc. These features aren't likely to show up in EIDE drives simply because adding them would cost the same as SCSI and why reinvent the wheel?

Right now the next step in consumer high-speed drives appears to be Firewire/iLink/1394 (depending on the vendor.) USB 2.0 has just appreared in silicon but it's already slower then Firewire/iLink/1394 and not as flexible. Intel is also working on PCI/X as a next generation replacement for the now venerable PCI bus. They appear to be going to a serial-bus design with smart interconnects.

One well regarded scenario for the future of PC's has them turning into black boxes containing little more then a CPU and graphics card. Everything else would be handled through high-speed serial connections.

Re:SCSI & Multi-tasking OS's

[Tower]

There are any number of vendors that are working on PCI-X, but PCs would still be better served if we could see some 64bit slots in them. Any 33MHz card will cause the 66MHz(PCI) or 133MHz(PCI-X) bus to operate at 33MHz (unless you add more bridges so they are separate phsical busses), but 64b slots can buy you a lot with not too much effort (a bunch of extra board traces, chip pinout, connector, etc) and not all that much more cost. Heck the DEC Alpha PC boards (note: 21164PC chip, not PC as in x86) had 64b PCI since, what, 96?
......
--

Re:cause the wheel is broken

[OnyxRaven]

ah but you forget - SCSI 4 and 5 are on the way - up to something like 480mb/s.

Re:cause the wheel is broken

[Tower]

Ultra2/80 performs remarkably better than ATA/100, and supports a load more devices, too... Ultra 320 (packetized SCSI - like FC) is not too far off, with 640 being just a little pipe-dream at the moment (the silicon tech required for the SCSI drivers... well, that many lines with that much speed with that much driving capability all on one chip is a tall order).

Once ATA gets fixed (trashed) we won't need to bother with it anymore... damn legacy 8^)

--

4x the cost? What do you smoke?

[Wakko Warner]

Western Digital WDE18300-0048 18.0GB ULTRA2 WIDE SCSI 68pin 7200RPM 6.9MS 2MB CACHE 18.2GB $243

Western Digital18.0GB EIDE 7200 RPM OEM,1 YEAR WARRANTY $109

Doesn't look like 4-to-1 to me. I'll pay the extra $130 for SCSI, personally.

- A.P.
--

"One World, one Web, one Program" - Microsoft promotional ad

Re:cause the wheel is broken

[Detritus]

Besides the advantages that have already been mentioned, Ultra160 supports much longer cables (12 meters) than any variety of IDE. Assuming you had enough IDE controller chips, how are you going to connect 8 hard drives to a PC with 18" cables?

Re:4x the cost? What do you smoke?

[Wakko Warner]

Or, do what I do, and buy IDE for mass "garbage" storage and SCSI (soon RAID-ed) for speed.

You get what you pay for.

- A.P.
--

"One World, one Web, one Program" - Microsoft promotional ad

Re:SCSI encapsulation for IDE drives: found

[Matt_Bennett]

It has already been done, at least as a RAID solution. I was searching for fast, huge storage solutions 3 years ago and found a company that was making SCSI RAID drives based on IDE. I can't remember the company that I talked to back then, but here's one link I could find with a brief search.

I think if you take it to RAID-0 you can array all those IDE drives to look like one HUGE SCSI drive.

Re:SCSI encapsulation for IDE drives

[Xenu]

Many SCSI drives used to actually be ST506 and ESDI drives with SCSI bridge boards installed by the manufacturer. There is no reason that it couldn't be done with IDE drives.

3Ware's ATA-RAID controllers do just that!

[BitMan]

Forget Promise, SIIG and others. 3Ware's Escalade series of products are just what you are looking for. Keys to performance with Escalade:

• On-board co-processor that acts like a SCSI target from the standpoint of the OS/driver. The same you'll find on most SCSI RAID controllers (i960 or similiar). This dedicated CPU drives your CPU, not your mainboard chipset's southbridge (which normally requires some CPU overhead even with bus mastering).
• One IDE drive per-channel. No "slave" issues. 100% Hot-Swap capability (although you'll need a IDE hot-swap bay/chassis for full hot-swap capability). Maximum performance.
• 2-8 channel boards, roughly $50-60/channel -- not much more than those crappy Promise FastTrak cards, only much, much faster.
• 100% Linux support. 3Ware controller support is built-in to most newer 2.2.x kernels.

If you want to minimize cost and performance, 3Ware's Escalade is what you want. Their new 6000-series offers 2-8 channels of RAID-0/1/1+0 with Ultra66 support for $139/279/479 (2/4/8 channel).

3Ware is also working on a 64-bit PCI board with RAID-5 support (as well as Ultra100). Be looking out for it (I know I will).

-- Bryan "TheBS" Smith

Re:3Ware's ATA-RAID controllers do just that!

[BitMan]

Er, that should read "dedicated CPU drives your disks" not "dedicated CPU drives your CPU." [ DOOH! ]

Also wanted to point out that with any striping, mirroring, parity, you are "bothering" your CPU with BIOS or software RAID solutions (and the Promise solution *IS* a "BIOS" RAID solution). The 3Ware controller will off-load these routines onto it's own co-processor.

-- Bryan "TheBS" Smith