XP RAID vs md
Wow... I hope that one of the maintainers will comment on this, I didn't even know that XP had a sw RAID implementation. Up to 100% more on read and 60% more on write is quite a significant margin. Is there anything in favour of the md driver if this is true?
Quoted 4 lines · Vladimir Milovanovic
Wow... I hope that one of the maintainers will comment on this, I didn't even know that XP had a sw RAID implementation. Up to 100% more on read and 60% more on write is quite a significant margin. Is there anything in favour of the md driver if this is true?
I've not really checked these numbers yet, so take the following with a grain of salt.
However, with RAID1 for example, I got approximately twice the read speed and 95% of the write speed (compared to just using a single disk).
I have a really hard time imagining a 100% read boost; that would simply exceed disk bandwidth, and 60% writes - how should that work?
I'm not claiming md is perfect or the fastest imaginable solution, but it is rather close to theoretical disk bandwidth. A two digit percentage performance improvement just can't be done.
Sincerely,
Lars Marowsky-Brée <[email protected]>
Signature · 8 lines
SuSE Labs - Research & Development, SuSE Linux AG "If anything can go wrong, it will." "Chance favors the prepared (mind)." -- Capt. Edward A. Murphy -- Louis Pasteur - To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Lars Marowsky-Bree on Wed 2/04 17:27 +0200:
Quoted 11 lines
Quoted 5 lines
Wow... I hope that one of the maintainers will comment on this, I didn't even know that XP had a sw RAID implementation. Up to 100% more on read and 60% more on write is quite a significant margin. Is there anything in favour of the md driver if this is true?
However, with RAID1 for example, I got approximately twice the read speed [...]
I have a really hard time imagining a 100% read boost; that would simply exceed disk bandwidth [...]
isn't "100%" the same as double? If you get "approximately twice" then you are getting 100% already...
[posted & mailed]
Quoted 7 lines · Lars Marowsky-Bree
Quoted 4 lines
Wow... I hope that one of the maintainers will comment on this, I didn't even know that XP had a sw RAID implementation. Up to 100% more on read and 60% more on write is quite a significant margin. Is there anything in favour of the md driver if this is true?
I've not really checked these numbers yet, so take the following with a grain of salt.
Have a look at http://www.hpl.hp.com/techreports/2002/HPL-2002-352.pdf
- that'll tell you were the figures are from - the 100% is worst-case.
Quoted 2 lines
However, with RAID1 for example, I got approximately twice the read speed and 95% of the write speed (compared to just using a single disk).
One thing I forgot to mention in my post is that this is all (AFAICS) RAID0.
Quoted 2 lines
I have a really hard time imagining a 100% read boost; that would simply exceed disk bandwidth, and 60% writes - how should that work?
The md driver seriously underperforms at certain request sizes, and is generally underperforming. The average boost is less, but still about 40-60% read and 30-40% write. XP's RAID0 is better than JBOD above a certain request size, which is pretty good.
Quoted 3 lines
I'm not claiming md is perfect or the fastest imaginable solution, but it is rather close to theoretical disk bandwidth. A two digit percentage performance improvement just can't be done.
Well, that's why I asked in my original post for people to look at the report. Is isn't that long and it's pretty clear (although short on config details).
Signature · 4 lines
Illtud Daniel [email protected] Uwch Ddadansoddwr Systemau Senior Systems Analyst Llyfrgell Genedlaethol Cymru National Library of Wales Yn siarad drosof fy hun, nid LlGC - Speaking personally, not for NLW
Well Daniel what is your take on this? I am very interested in this since I am working on a project which has to use RAID, and performance is paramount (after reliability of course). I need to make a decision, and I would like very much to stick with md, since it is what I know and am used to. But if one can get huge performance margins using XP, then I am going to have to consider that.
Cheers,
Vlad.
Quoted 66 lines · Illtud Daniel
[posted & mailed]
Quoted 10 lines · Lars Marowsky-Bree
Quoted 6 lines
Wow... I hope that one of the maintainers will comment on this, I didn't even know that XP had a sw RAID implementation. Up to 100% more on read and 60% more on write is quite a significant margin. Is there anything in favour of the md driver if this is true?
I've not really checked these numbers yet, so take the following with a grain of salt.
Have a look at http://www.hpl.hp.com/techreports/2002/HPL-2002-352.pdf
- that'll tell you were the figures are from - the 100% is worst-case.Quoted 4 lines
However, with RAID1 for example, I got approximately twice the read speed and 95% of the write speed (compared to just using a single disk).
One thing I forgot to mention in my post is that this is all (AFAICS) RAID0.
Quoted 4 lines
I have a really hard time imagining a 100% read boost; that would simply exceed disk bandwidth, and 60% writes - how should that work?
The md driver seriously underperforms at certain request sizes, and is generally underperforming. The average boost is less, but still about 40-60% read and 30-40% write. XP's RAID0 is better than JBOD above a certain request size, which is pretty good.
Quoted 5 lines
I'm not claiming md is perfect or the fastest imaginable solution, but it is rather close to theoretical disk bandwidth. A two digit percentage performance improvement just can't be done.
Well, that's why I asked in my original post for people to look at the report. Is isn't that long and it's pretty clear (although short on config details).
--
Illtud Daniel [email protected]
Uwch Ddadansoddwr Systemau Senior Systems Analyst
Llyfrgell Genedlaethol Cymru National Library of Wales
Yn siarad drosof fy hun, nid LlGC - Speaking personally, not for NLW
-
To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 3 lines · Vladimir Milovanovic
Well Daniel what is your take on this? I am very interested in this since I am working on a project which has to use RAID, and performance is paramount (after reliability of course).
...in which case I'd go for a hardware SCSI solution. The AMI Megaraid (now LSI?) cards have served me well. You can afford one by not buying XP...
Quoted 4 lines
I need to make a decision,
and I would like very much to stick with md, since it is what I know and am used to. But if one can get huge performance margins using XP, then I am going to have to consider that.
You must remember that this is RAID0, and says nothing about other RAID levels. I'm assuming you wouldn't be implementing a RAID0 (not on its own, anyway - maybe 5+0) if you're looking for reliability. You've also got to look at where you need the speed - is this a network file server? In which case your bottleneck is probably the network, unless you've got more than two gigabit ethernet cards. Tell us more about what your project needs RAID for, and we can (probably) tell you that md is just what you need. I'd be very wary of going over to XP on the back of this one report, since you're opening yourself up to a load of extra costs (on everything!). If you've got the money, then why not just go for hardware SCSI RAID?
I've benchmarked NT4.0's RAID0 of two hardware RAID 5 arrays against linux's (about a year ago) and linux smoked NT, which is why this report surprised me, I wasn't aware that XP had come on so far, and I was after some feedback from the md developers as to whether they could spot another reason for the results in the report.
Signature · 4 lines
Illtud Daniel [email protected] Uwch Ddadansoddwr Systemau Senior Systems Analyst Llyfrgell Genedlaethol Cymru National Library of Wales Yn siarad drosof fy hun, nid LlGC - Speaking personally, not for NLW