From Karl-Heinz Herrmann
[Heather] The question this time is from our own Karl-Heinz Herrmann. When you play on the edge, it's best to ask a few buddies to keep an eye out for the loose rocks...
Answered By: Robos, Rick Moen, Heather Stern
I'm hunting for new desktop hardware and getting confused and somewhat frustrated.
[Robos] Me too (sort of)
The best German computer magzine (c't) has of course plenty of tests and recommendations as well as "optimal PCs" for certain usage profiles. My problem is, I want a decent performing raid 5 (burned to often by sudden drive failure).
[Robos] Have that already in the gangway - working flawlessly for several years now. 80GB raid 5.
[Rick] Do you expect to be so incredibly short of disk space that RAID1 isn't an option?
[Heather] Funny you should mention that; I met someone recently who got burned by sudden drive failure on a RAID 1 system - and to their horror, while the other drive was fine contents-wise, it was also in a state where it couldn't be brought up normally to be looked at - expected a twin to be present, and nobody wanted to risk it considering a "mew" drive the master and wiping it. It needed a visit to the drive recovery shop.
RAID is not a panacea - a cure-all - against hardware failure; it merely helps. Usually, a lot. But it's specifically not a solution against slow corruption via software failure. Anyone working with oversized disks should seriously consider their backup-and-recovery plans too.
no -- raid1 would be an option. If some 50GB actual space are there that should be enough for some time, (temp) video data won't need raid1.
With recent consumer boards I seem to have the option of using IDE drives, sure. Cheap huge drive space and all. If I understand the board details right all of the regular ATA interfaces are on the PCI bus -- so a raid with 3 or 4 drives there is hogging the PCI bus. There are SATA interfaces which are not going through the PCI interface but have a separate one -- just Linux seems not to happy with SATA yet and most boards have 2 SATA drives -- not enough for raid5. If I could be sure that a raid1 is running stable and at full speed in Linux with SATA drives hat might be an optin, as SATA brings plenty of disk space for a reasonable price).
[Rick] Huh? RAID1 will inherently be faster than RAID5. But either is fine with the md driver.
That's not what I meant -- from all I heard SATA is still highly experimental and often not supported at all. Most often I heard as problemsolver to switch them to IDE.
[Rick Moen] That's one workaround, of three total.
I'm not quite sure if the SATA drives are connected to the IDE connections for that or if the SATA connections can be used in a compatibility mode.
[Rick] The latter -- where the BIOS supports doing so.
And I also don't know what performance changes that might bring along.
[Rick] Who cares? You use the drive in that mode only long enough to compile a new kernel with the requisite support for the SATA chipset in question.
I've been googling quite some time again and slowly I get an idea on Linux support. libata is providing sata support --
[Rick] More precisely, libata is providing better and broader SATA support.
patch in 2.4, already included in 2.6 but still under heavy development as the latest changelogs and bugfixes show.
[Rick] Correct. However, there is non-libata support for several SATA chipsets in recent 2.4 kernels. (Of those, 3Ware Escalade is supported for a long ways back in 2.4.x.)
I'm not sure about the Via K8T800 chipset
[Rick] That isn't a SATA chipset. It's a motherboard chipset -- always accompanied, to the best of my knowledge, by the VIA VT8237 South Bridge, which does SATA and is supported in libata.
but the nforce3 seems mostly supported (sound seems a dead weight).
[Rick] Nvidia nforce3 is also not an SATA chipset, but rather a motherboard one -- always accompanied, to the best of my knowledge, by a really awful Promise R20378 SATA / software-RAID chip. The sound is indeed provided by a really dreadful Realtek ALC650 chip.
I would run, not walk, in just about any direction leading away from that kit.
There are driver tar packages from nvidia for download (at least partly source, not sure if some binary only parts are in there too). The Sata controller on nforce3 boards seem to be Silicon Image like Sil3112A, Sil3114.
Robos: You said they are supported, do you have pointers for details?
[Rick] That is conceivable, but is not what I've been seeing. It's possible that it's dubious-quality Silicon Image chips in some cases, and awful quality Promise chips in others.
[Robos] Yes, my flatmates experience But he's on holiday so I can't ask him. And I don't really recall any details.
[Heather] I've got one; a buddy just compiled 2.6.1-mm4 and with a couple of stumbles making sure other parts work also, is now ready to fall the rest of the way in love with his Athlon/SATA system, bearing an sil3114 chipset. Statistics on exactly how cool not available just yet
Has anyboy here SATA drives working? Performance? kernel versions? any patches? And on what particular chipset?
[Rick] Best quality: Adaptec 24x0, 3Ware Escalade -- in that order. But you won't find those built into motherboards. Usually, what you find inside motherboard chipsets is Intel ICH5. Which is basically OK.
I don't mind AMD/P4 or even AMD64 (new 3000+ is reasonably priced).
[Robos] OK, I don't have sata drives but I read something about it. It seems (see latest c't) that all the upcoming athlon 64 boards have sata (since it's in their north|southbride) and quite a lot of them already have hardware raid0 and 1.
[Rick] You can read more about it here: "Serial ATA" on http://linuxmafia.com/kb/Hardware
Yes -- the hardware is there -- but as far as I know none of the hardware raids are supported in Linux, and booting Linux off SATA drives which run actually as SATA and not in some IDE-emulation mode is difficult. What I do not know if once a Lniux kernel takes over if they are running as real SATA? Performance penalty for IDE emulation mode? And the boards tested in latest c't all have two SATA connectors, i.e. to few for raid5. Of course I could plug in 1 or 2 additional SCSI disks on the PCI bus with them.
Alternatively I've already plenty of SCSI hardware in the old one, so I would maybe even prefer a decent SCSI setup -- but I can't find a chipset with onboard SCSI controllers -- and if I plug PCI cards I'm again hogging the PCI bus.
Are there any other chipsets besides nforce2/3, i875,....?
Gods above, Karl-Heinz, what sort of incredibly disk-intensive operations are you planning to conduct?
Well -- right now I've some 32% CPU at full whopping 9MB/s (bonnie 8892 kB/s 31.8 %CPU) from the SCSI raid5 in a Pent. classic board running a K6II-366MHz. Since I am hunting for new hardware I am thinking of getting something decent which I (again) can keep for some time. The old system is running since 1997 after all, with occasional low cost upgrades(like the K6).
My original idea was: Ok upgrade is necessary, IDE is still not what I want (performance trouble with my existing Lap and a DVD burner in the desktop put me off; that cdrecord mess, ide-scsi or not, DMA only with the right block size,...). So how about some new SCSI drives (3 for raid5) and a nice new card (U160 one channel should do as I still have a 2940UW for CD, scanner, etc.) and be done. Well-- the lowest grade (new) SCSI drives I can locate are some 10k rpm 18 or 36GB drives which sustain easily 30+MB/s and peak at 70 MB/s. A PCI bus is getting very much the limiting factor as there are probably more cards in there as well and I'm looking for a way to avoid building a bottleneck into the new thing right from the start. Thats why I tried to ask for info's on boards/chipsets which might have SCSI onboard, preferrably circumventing the PCI bus.
But you are probably right, just ignore the bottleneck and let 100MB/s be good enough. In a year the SATA support will be there and offers a non-PCI extension anyway if needed.
[Rick] The SATA 1.0 spec claims a 150MB/sec limit on bus transfers. Of course, the fastest ATA hard drive manufactured today can, under contrived, artificially ideal conditions, read data at maybe 50 MB/sec, so saturating ATA/100 or ATA/133 sounds very farfetched for a long time to come. (Remember, only one ATA device per bus can be active at any specific time. SATA changes this only by permitting connection of only one ATA device on the chain.
Aren't there any "server-boards" with other specifications? In the price lists here none show up, but a recent acer ad mentioned some "ServerWorks Chipset GC-SL".
[Robos] IIRC Silicon Image chips were good under linux. One of the kernel developers even managed to persuade some hw producer to donate some driver to gpl. Can't recall which one that was, might have been promise.
Any boards/manufacturers using that chipsets (available in Germany)?
[Robos] OK, I just looked on alternate.de and didn't find a silicon image but rather found this: http://www1.alternate.de/html/nodes_info/giea01.html
- Regarding their promise chipset I found this:
Can anybody point me to details what this chipset can do (http://www.acer.com doesn't)? Or knows of a nice one (like onboard SCSI [raid] controller bypassing PCI, don't forget the Linux support)? Or at least 64 bit PCI? All the consumer boards showing up in a A4 page (tiny print) of boards seem not to have anything special in that regard.
Am I overemphasising the PCI-hogging? To cite above mentioned c't -- If you run a Gigabit ethernet on PCI it basically saturates the PCI bus.
[Rick] 1. Yes.
[Rick] 2. Using Linux software RAID (md driver) for redundancy, you typically won't even notice the load relative to (say) a hugely expensive Mylex SCSI hardware RAID adapter except during RAID volume rebuilds (if you ever have to do that). In such a case, expect the system to very, very seriously bog down during the rebuild. With the Mylex card, all of that would be handled entirely within the array, with basically no system loading.
Every unclean reboot -- but the raid5 regeneration isn't that bad, even on the old system.
[Rick] I'm talking about when one replaces a failed ATA drive in the array. System will be usable but heavily loaded.
Takes some 10 minutes each 3-5 GB mdX. System is still usable during that time. Rebuilding is on a low prioritiy.
I picked up another thing while reading through kerneltrap mailinglists: It seems most onboard so called IDE raid cards are actually just fancy labelled IDE adapters where the windows driver is nothing but a software raid.
[Rick] Yes. Avoid. Use the md driver, instead.
So raid with IDE raid cards are still software raid unless its the (rather expensive?) real thing.
[Rick] Yes. "Real thing" basically means 3Ware Escalade 850x or Adaptec AAR 24x0.
Promise was explicitly mentioned.
What I'm not sure is if this is for all raid modes or just raid 5.
[Rick] Please see my file. You can sometimes get RAID0 disk-spanning using a BIOS trick.
Simply doubling the writes for a raid 1 can't be that demanding in card logic after all.... Does anybody know more on this? Has anybody used IDE raid controllers? How do they behave?
[Rick] 3. Indeed, gigabit ethernet will strain practically any system, even with jumbo MTU sizes set, and even with a suitable switch and other associated equipment.
So that's true -- but then I've no immediate use for a Gbit ethernet now but I want to keep the option open. I think I'll watch the AMD64 scene for some more time, especially if hardware drivers for something like TV cards will run with 64bit OS.
If all this bleeding edge driver location stuff is getting sufficiently annoying I might even grab some PIII off ebay and just plug my existing SCSI chains in (and swap some 9GB drives in for the 2 and 4.5 one)
Still -- specs on that AMD64 sound nice.
[Rick] Yes, I certainly am with you, there! But I want to jump onto the 2nd generation of Opterons, not the first.
There might be a fair bit of video processing in the near future (as well as the use as home media center with timeshifting and all that stuff) so there will be incoming video streams on the PCI as well.
[Robos] we use a 400Mhz Celi for the (software) raid 5 and it works like a charm. Sometimes my flat-mate and I both stream stuff from there and no clogging or anything. Regarding timeshift: if you use dvb you can easily capture the stuff since it's already in mpeg2 format. No processor usage then. And any recent platter should be able to both record and play mpeg2 streams simultaneously.
I am toying with the idea of buying an athlon 3000+ soon: no intel (they have a too high profit margin for my liking), very performant, rather cheap, 64bit (for what it's worth) which linux (debian IIRC) supports rather good. But stay away from the vobis offering, that sucks
That AMD3000+ is intriguing -- what a small c't article can do to ones minds
[Robos] There was one before, like 5 numbers or so, which dealt with 64bit again 32 bit. You might want to read that before too.
|Meet the Gang 1 2 3 4|