Serial ATA and Serial SCSI

aibrahim writes "In the recent Slashdot article about Serial ATA some people wanted to know where SCSI was going, and if Serial ATA could deal with some higher end workstation and low end server requirements. Apparently it has been decided that Serial ATA 2 (pdf doc) and Serial Attached SCSI are the answers."

Comments

Copy/Paste a previously published article, anyone?

[jhoffoss]

Nearly anything is better than ribbon cables. aibrahim writes "In the recent /. article about Serial ATA some people wanted to know where SCSI was going, and if Serial ATA could deal with some higher end workstation and low end server requirements. Apparently it has been decided that Serial ATA 2 (pdf doc) and Serial Attached SCSI are the answers."

http://slashdot.org/article.pl?sid=02/07/04/153224 =nested=167

Difference

[SpatchMonkey]

The main difference is that Serial ATA will be more readily available first, and will therefore become more popular.

If you look at the Serial SCSI page in the FAQ, note that it is still under development, where motherboards supporting Serial ATA are out now.

Sheesh!

[Mike Hicks]

...and just as I'm building my fibre channel array for my home computer :-p

Where are the drives?

[charnov]

Okay...I have a SATA equipped mobo on order which comes in in two weeks. What I want to know is: where are the hard drives? And no, I don't mean the drives that are standard ATA100 that have converters. I mean Seagates native SATA drives. They demoed supposed "production" pieces back in Fenruary.

Re:Where are the drives?

[bjschrock]

The Segate Baracudda ATA V, Serial ATA version is going to be released "this fall" according to a press release I read somewhere. It has a 8MB cache and comes in sizes up to 120GB.

Re:Where are the drives?

[MrResistor]

You can bet the first SATA drives are going to be quite a bit more expensive than an equivalent capacity ATA100, too.

Not according to Seagate. I wish I could remember where I read that, but it was fairly recent.

Serial SCSI is neat.

[Sheetrock]

I've read a couple of brochures on this technology (we're looking at high-bandwidth high-availability file clusters for our hybrid AS/400-Solaris data warehousing) and it looks extremely promising.

Basically, they're extending parallel SCSI technology to address next generation I/O and direct attach storage requirements. It uses the (proven) interface from Serial ATA to avoid an unnecessarily proprietary interface and the costs that usually entails. The naming is unfortunate, because one usually thinks of parallel (side-by-side) as being faster than serial (one after the other) when the technology allows you to combine the two tactics much like in LANs. This is the technology that will enable a new generation of dense devices, such as small form factor hard drives, whereas Parallel SCSI can't because of cabling and voltage issues.

So depending on the pricing of the technology when it hits the shelves/junk mail catalogs, we're going to take a serious look at it. Does anybody have any prototype benchmarks?

I believe..

[iONiUM]

I believe the general conception that SCSI is too expensive for the home user is going to make it hard for SCSI to pick up now. Although SCSI is much faster (and better for business in my opinion), I think IDE will continue to rule with it's slower perfomance and cheaper prices the home market. Quite a shame though, IDE seems to always be so slow when compared to the incredibly fast SCSI drives out... but then again, the size of IDE w/ current prices means you can get a huge hard drive for relatively cheap, which is almost impossible with SCSI.

Re:I believe..

[Courageous]

Well, you do know, don't you, that there is no reason why Seagate can't be selling those 15,000 RPM Cheetahs with an IDE interface if they wanted to, right? The SCSI bus issue doesn't significantly impact the drive speed in any way. It's a _BUS_.

C//

Re:I believe..

[robertchin]

The general consensus is that SCSI is faster, but in reality, SCSI is only faster when there's more than one hard drive per bus. If you have only one drive per bus, IDE is faster than SCSI. So basically if chipsets supported four IDE buses so that you could separate each drive onto its own bus, SCSI would really not have any chance of being faster. This is what apple does on its XServe, give it a distinct price/performance advantage. If you really want speed, you go to FC-AL.

Re:firewire

[bofkentucky]

Lets look at it realistically:

• 800Megabits/s = 100MegaBytes/s, so transfer rate is the same as a standard IDE disk, after the speed bump.
• What's apple charging for a firewire license these days? Will impact pricing of controllers and drives?
• You mention hot-swapability, but most ATA drives aren't in hot swap capable enclosures.
• Is it just the interface or do other things need to change to allow hot swaping (More drive components = more expensive)
• Are there RAID 1+0 controllers for firewire drives?

Re:OSS...

[SirSlud]

I dont get it. Are these, like, 40 year old project manager trolls? Or, like, 19 year old scriptkiddie "Nike makes the best shoes, MS makes the best software" trolls?

I wanna know the demographic profile here that for some reason feels threatened by Open Source. Pure curiosity.

just like the RAMBUS story

[lingqi]

remember back when RAMBUS said: we will provide an architecture with very narrow bus but extremely high speed to make up for it? (the *original* RAMBUS specs) -- beside the royalties and whatnot -- it actually (technology discussion only) had merits in that the PCB design was greatly simplified because of less crosstalk, easier routing, etc etc.

and then, people demanded more bandwidth... so now we have double / quad pumped RAMBUS channels -- in the end (today) it's back to 64-bit data-bus *anyhow*... except with an architecture that's not designed for parallel operation.

do anybody see some parallel (ha!) here?

i am guessing (or, predicting) that serial ATA / SCSI will go the same route. i really hope that it won't -- because if it did, our lives will all be kinda rough -- but it probabbly will.

sigh...

Re:just like the RAMBUS story

[jamesbulman]

Double / Quad pumped rambus channels refer to the frequency at which they are being driven, not the number of traces laid down. Rambus is as narrow as it has ever been.

Re:just like the RAMBUS story

[cheezedawg]

Ethernet has had no problems scaling to higher bandwith while maintaining its serial "bus". Serial ATA is a packetized interface that is more similar to ethernet than RAMBUS. They already have 600 MB/sec SATA on the roadmap.

Re:Serial ATA v. SAS

[iKitty]

Your performance assumptions are flawed. U320 SCSI allows 320MB/sec for the entire SHARED SCSI BUS (15 drives). S-ATA gives you 150MB/sec to EACH drive; the SATA controllers support a Direct Port Access (DPA) mode that allows DMA transactions to proceed simultaneously to all drives. With the new Seagate and WD drives having the larger 8MB cache in each drive we will see decent per drive performance and in large drive arrays (>8 drives) the performance should beat the pants off of SCSI.

Evil Drives

[Blahbbs]

So when will someone come out with a Serial ATA Network device?

Yup, a SATAN Device (tm).

...boo...hiss....

I heard

[Apreche]

that serial ATA, while being very fast and much better than what I've got now, will have DRM built in. Is this true? Should I not get serial ATA in my next system because of it? Anyone got any links pertaining to this issue?

Re:Internal firewire?

[alienmole]

Current Firewire is half the bandwidth of ATA/100. Theoretically, Firewire can be saturated by a single fast disk, so for internal Firewire you'd want a separate Firewire channel on the motherboard for each disk - but Firewire was really intended to be a serial bus which supports multiple devices. That's why in its current form, Firewire is more appropriate for connecting devices like video recorders, or hooking up a single external drive for data portability, than for internal drives.

Re:firewire

[alienmole]

You're confusing different theoretical limits. The theoretical limit of current Firewire is 400Mbps. The idea is that the Firewire spec allows for future versions of Firewire to reach 1600Mbps, but current Firewire buses aren't even theoretically capable of that. In practice, 400Mbps Firewire doesn't get much above 315Mbps of actual data throughput - consistently slower even than ATA/66, which has been demonstrated in real-world tests.

OTOH, ATA/100 is theoretically capable of 100MB/s, i.e. 800Mbps, in its current incarnation today. You probably won't hit that in practice, not only because of bus limitations, but because 50MB/s is about the max any IDE drive can pump out, so you'd need two drives running continuously at their absolute peak speed in order to saturate the ATA/100 bus.

Re:nope

[Fulcrum of Evil]

what i am refering to is "dual-channel" RDRAM.

The thing to remember there is that you don't have to synchronize the channels - they go as fast as they go. Multi-channel interleaved ram is a pretty easy way to speed up access, but it costs more (of course)

ATA/SCSI distinction

[XNormal]

With current silicon integration levels there is no real reason why SCSI should be more expensive than ATA. They could have just merged them and perhaps emulated braindead ATA on top of SCSI to keep compatibility or something if anyone really wants to.

I'm pretty sure the only reason they keep the difference is to be able to charge more from people building servers. It's purely a marketing and price positioning decision.

Re:i think you are getting confused

[cheezedawg]

No, my point was they didn't have to add more signals to increase the speed of ethernet, and they probably won't have to with SATA either. Look at all of the hot technologies now: USB(2), firewire, 3GIO, SATA, and Serian SCSI. They are all serial interfaces.

we might as well just keep on using parallel ATA but boost the signal freq incrementally, since it will get us to the same place in a few years anyway

One of the biggest reasons for developing SATA was that parallel ATA is pretty much maxed out at 133 right now- we can't just "boost the signal freq" any more. Plus, parallel ATA is based on TTL signaling, and that requires the integrated circuits to tolerate +5V input signals. This is getting harder and harder to support with the modern manufacturing processes of the chip. And as you pointed out, the fewer signals also has the benefit of simplified design and reduced the cost of the chips.

by the way -- PCI bus can only push 133MB anyhow -- anything beyond that is silly

I hope you weren't serious about that.
#1. If PCI is becoming the bottleneck, then we will move past it. In fact, we already are. PCI's replacement (3GIO) is already in development (actually I think they changed the name to PCI Express- kind of dumb if you ask me).
#2. Integrated SATA implementations will not be on the PCI bus, so they will not be limited to the bandwidth of the bus. They will only be limited by the upstream bandwidth between the southbridge and northbridge chipsets (on current Intel desktop chipsets, this is 266 MB/sec with plans to increase to 533 MB/sec soon- AMD's Hypertransport is also more than adequate for SATA)

So?

[gerardrj]

They still don't say that serial ATA will support more than two devices per channel. In fact they say it will be software compatible with ATA in its current form, suggesting it continues the master/slave relationship.

Today's drive media can only reach 40MB/s reading from the platters for short bursts, if their lucky. Normally they'll read/write about 20MB/s. What's the point of another boost in speed of ATA (to the suggested 150MB/s) when you will only ever be able to use 80MB/s of that. Oh, that's right... the ignorant users need bigger numbers on their cardboard boxes to show off to the neighbors.

Does anyone have any information on a HD soon to be released that will offer a quantum leap of read-from-meadia performance to something like 75MB/s? That's more than triple the current read-from-meadia speeds, and they seem to only ever increase the speeds by about 1-2MB/s each year.

SCSI makes sense having very high bus bandwidth, as you can connect quite a few devices and use the connect/disconnect to send simultaneous reads/writes to multiple devices. In that scheme, you can keep most of your drives operating at the same time. Of course Apple has shown that at least for a small RAID, multiple independent ATA channels are just as fast and lower cost than a single SCSI channel. I persoanally have a difficult time thinking that multi-ATA design would scale well to a 32 drive RAID, where a dual channel SCSI would shine.

Re:So?

[gerardrj]

But the buffers are insignificant. an 8MB buffer will be emptied or filled over the wire in .06 seconds at 133MB/s. So you can't get max throughput for anything longer than about .2 seconds.

Since the caches on the drives don't undstand filesystems or file structure, they can only contain things that have already been read from disk, or assume the next read will be a sequential block and pre-fetch that. More often than not, the cache on the drive does not contain the requested data. The disk cache only helps for small files that are re-read often like directories, and really the OS's disk cache will provide even better performance in these situations. The drive buffer does nothing to increase real-world data throughput on ATA disks, it's just there so the drive makers can claim a really high (wire speed) peak throughput number. Caches do make sense on SCSI drives where the drive can be ordered to read a set of blocks to buffer, disconnect, and later have the blocks read from buffer. During that drive's read phase (while disconnected) other drives can be commanded to read or write data to/from their buffers. This is why SCSI RAID systems outperform ATA RAID systems.

As for the increase in drive throughput from media: if future advances play out the way the industry has advanced in the past, it will be 15-20 years before a drive will be able to move 100MB/s sustained from rotational media. 10 years ago we where getting 10MB/s sustained, today we are getting 20, sometimes 30. Switching to some non-rotational media might see throughput increase dramatically, but all such devices I've seen connect to Firewire or USB[2] thus negating the need for more ATA bandwidth.

Serial ATA is a project in search of a problem, or perhaps more accurately marketing hype in search of consumer dollars.

Re:So?

[pangloss]

ok, i'll bite:

But the buffers are insignificant.

this is just theory, right? or do you have some benchmarks that you can link to?

i seem to recall that storagereview showed some marked performance improvements between the western digital xxxxBB and xxxxJB (aka "special edition") drives that supposedly only differ in the amount of cache (2MB vs 8MB).

Re:So?

[gerardrj]

I'm looking at a page on storagereview.com regarding the BB and JB models.
Summary:
Xfer rates: outer( BB: 49.3 JB:49.0) inner( BB:29.2 JB:29.2)
They don't mention the type of data they are writing, but it appears to be sequential reads/writes, probably in sector sizes. In any case the numbers are essentially identical. Looking at the 'desktop performance' page, there are wider, but still rather insignificant differences in performance. Certainly small enough that other factors could be causing some or all of the difference between drives.

Re:So?

[pangloss]

look at this one too, though. the numbers you were comparing are for two pretty different models--different capacities, platters, etc.

and here's a quote pulled from the above link:
"To differentiate their offerings from the competition, one of WD's largest OEMs recently requested an ATA drive with an 8-megabyte buffer. The manufacturer responded by retooling its current flagship, the WD1000BB, with an 8-meg cache."

which suggests that the cache bump was the only difference.

Re:So?

[gerardrj]

Okay... that result also shows almost no difference between the different models.

Serial SCSI?

[ghopper]

Isn't that called Firewire?

SCSI vs Firewire

Re:i think you are getting confused

[cheezedawg]

FWIW, the major OEM's have been clamoring for SATA because the SATA cables are cheaper than PATA ribbon cables.

And the 70+ companies in the SATA Working Group, the PCI-Sig (for 3GIO), and the USB and Firewire designers disagree with your assessment of the scalability of these serial interfaces.

And replacing the 26 (or whatever) signal pins that are currently integrated into the southbridge chipset for parallel ATA with the 4 signals for SATA certainly does simplify MB design.

Re:Sounds like fun ...

[slaker]

No. Probably you don't. RAID5 parity calculations pretty much kill write performance on every array I've ever seen. RAID5 gives a lot of bang for your buck, but if you want for-real disk performance AND redundancy, RAID10 is the way to go.

The Problem with SCSCI Today...

[evilviper]

My complaint about SCSI is the fact that it's split among so many different implimentations. That means vendors are going to choose to support one, and ignore the rest.

What we need is a single Serial SCSI standard (woo, try saing that a few times). Instead of Fibre Channel, Serial SCSI, Firewire, iSCSI, and whatever else they've come up with, we need one single interface for them all.

If it wasn't for dirvergent implimentations (25/68/80 Pins-Wide/Fast SCSI) people would most likely have SCSI in their systems today, instead of IDE/ATA. ATA retained backwards compatibility, while SCSI gave up compatibility just to get to market with a slight speed boost, as soon as they possibly could.

The other problem with SCSI, configuration and addressing difficulties, will not be an issue with Serial SCSI.

So I'd be happy if I could buy a Serial SCSI card with front-mountable Firewire ports (with more bus power than current Firewire), and perhaps with an option to buy an adapter if I wish to hook that card up to Fibre Channel devices.

Remember, this criticizm is comming from someone who *HATES* IDE/USB.

The SCSI groups REALLY need to unite on this stuff if they want to see any sort of advancement, rather than each ending up as a niche technology. Just look a the Alpha systems. I'm convinced they could have easilly replaced x86 systems. If SCSI groups keep going this way, they face the same ultimate fate.

Re:The Problem with SCSCI Today...

[evilviper]

I can use an old scsi 2 drive on a scsi 3 system

Well, you pointed out several drawbacks yourself, and besides, IDE/ATA drives do not need adapters, and don't take a performance hit for that compatibility.

niche solutions are just fine for the needs of those niches

Not true. The larger the group using any technology, the less money each needs to pay. Besides, those outside the niche are getting screwed as well (getting stuck with ATA/USB), not to mention being force to get new perhiperals every time a new interface overtakes an old one.

Have you looked at Microsoft lately, it isn't the greatest but works-well-enough solutions, oh and it is cheap too?

Microsoft is not the best around (as SCSI is), is not faster and better than the competition, and certainly is not cheap. Hell! If anything, Microsoft is the ATA/USB of software. It sucks, but it comes with every computer you buy.

Re:Internal firewire?

[aibrahim]

IEEE 1394b supports bandwidth up to 3.2Gbps. TI has introduced a 1394b controller chipset.

Furtherless IEEE 1394b can run at 800Mbytes/sec over Cat 5 for up to 100m. Seems like you'll be able to just swap out your RJ-45 connectors for firewire ones and get to business.

As far as being saturated by a single fast disk...well do you have a single disk that can sustain 50Mbytes per second ? IEEE1394a can really transfer data very close to its theoretical limits in my experience. I've seen it shovel around 40+MBytes/sec so, I wouldn't write it off so fast. You need an IDE RAID 0 array to manage that.

Further, Apple is apparently considering rolling out 1394b as standard in the next round of desktops, and possibly the laptops too. (No link for the latter...)