Topic: Software RAID performances

[Walid Moghrabi]

Hi,

I just wanted to know how good (or bad) the performances with sofware mirroring is especially with huge (>60Gb) UWSCSI disks.

Is it comparable to simple disk usage or is it significantly slower ?
What are the requirement for a reasonnable usage knowing that I don't need FAST transfers but want good availability (it is aimed to be a shared directory for 30-40 users) ?

Thanks

[Chris O'Donovan]

I have twin 60GB Maxtors in one of my servers (1GHz Athlon) and hdparm measures almost 14MB/s with software mirroring. For the same non-raid disk hdparm measures 36MB/s.

Software mirroring seems to have a significant impact upon disk performance.

I suspect that filling the motherboard with RAM (eg, 768MB) would have a bigger performance impact than installing a hardware raid controller.

However, I don't think that this is the bottleneck in my server's performance. The maximum transfer rate I've observed with netatalk is 2-3MB/s and with samba 3-4MB/s. The CPU rarely goes past 2% utilization.

What are other people's experiences? Has anyone out there come anywhere close to the 10MB/s that fast ethernet is capable of with netatalk or samba? I have no problem getting 10MB/s with NFS.

Chris

[Walid Moghrabi]

Thanks for the information.
What about availability ?
I mean, I don't really care about transfert rate (I definitely never reaches 10 Mb/s in my network) but need not much latency because most of the files are accessed AND opened through my network so I don't want to much delay ...

Any informations about it ?

What about stability of E-Smith then ?
Is is something which can be considered as "robust" or is it too weak to be used in a 24h server ?

Thank you in advance ...

[Chris O'Donovan]

Latency? Is there an easy and meaningful way to measure this? The server I described above replaced an Apple solution and the users say that it is much faster than before.

Availability/Stability? Linux servers don't seem to crash. I have seen the netatalk daemon have problems when a client machine crashes while performing a network operation. Also, this client occasionally (every week or two) reboots the server because one of the user's sometimes has problems logging in. I suspect that the problem is with the user and not the hardware or software. Netatalk isn't as developed/supported as samba. The file locking, for example, seems to be much cruder.

Another installation which uses MSWindows for the clients doesn't have any problems. The uptimes on their servers are currently 16 and 31 days (which coincide with when the 5.0 upgrade was installed on each server). I haven't heard any comments from the users about the speed of these servers so I assume that they are comparable to the Microsoft servers they replaced (after a nimda virus invasion).

Because e-smith can be installed on almost any pentium-class machine I'd suggest installing it on an old machine you have lying around and use it yourself for a few weeks. Make sure that the machine has lots of RAM (a large disk cache makes a big difference to performance) and read the e-smith manual.

I suspect that you'll find that a 400MHz PII with 512MB RAM is more than adequate for 30-40 users and that UWSCSI isn't necessary. Please let us know your findings.

For a production machine a UPS and daily backups are vital.

Chris

[Filippo Carletti]

hdparm is not a good benchmark tool for raid
bonnie or bonnie++ should give more "real" results

Theory says that RAID1 has slower write performance and equal or slightly better read performance. I have verified theory on Linux md 0.90 many times.

Latest md has read load balancing, but I'm not sure it went also in redhat 2.2.19 kernel (sme5 kernel). Don't have time to check sorry.

And, please, put each hard disk on separate ide bus.

/dev/hda:
 Timing buffer-cache reads:   128 MB in  3.73 seconds = 34.32 MB/sec
 Timing buffered disk reads:  64 MB in 11.03 seconds =  5.80 MB/sec

/dev/hdc:
 Timing buffer-cache reads:   128 MB in  3.77 seconds = 33.95 MB/sec
 Timing buffered disk reads:  64 MB in 11.34 seconds =  5.64 MB/sec

/dev/md0:
 Timing buffer-cache reads:   128 MB in  3.55 seconds = 36.06 MB/sec
 Timing buffered disk reads:  64 MB in  8.31 seconds =  7.70 MB/sec

[Filippo Carletti]

> What about stability of E-Smith then ?
> Is is something which can be considered as "robust" or is it
> too weak to be used in a 24h server ?

Linux soft raid is stable.
Stable since 1999, I'd say.
I had one old mail server with kernel 2.2.6 and raid1 which stayed up for more than a year without any problem.