Another week, another DAOS implementation completed
Now, I'm not quite sure why IBM increased the default DAOS file size from 4K to 64K from 8.5.0 to 8.5.1 (the cynic in me screams early 8.5.0 benchmarks) but for this customer we changed it to 8K. I will say this again, RAID 5 is bad, RAID 5 is bad, RAID 5 is bad. If you are moving to 8.5.x and are looking at DAOS make sure you know what impact RAID 5 will have (in my tests RAID 5 is approx 20% worse than RAID 10). Also make sure to put the TX logs on a separate drive and controller if you can (at least do a separate drive).
This customer shaved about 52% overall on the storage required by Domino. Needless to say BRMS will also be a lot happier during daily, now somewhat incremental, backups.
Discussion for this entry is now closed.
Comments (4)
@1, well I broke my own rule. Never post about RAID levels, religion, politics, anti virus providers or ugly babies.....
My dislike from RAID 5 comes with many, many years of fighting slow write I/O performance. In essence, every single write (and read if last accessed it turned on and most are) will be written to every disk in the array. So if I have a 16 disk RAID 5 array, every single write operation is written 16 times. To say this is bad for performance is an understatement. My estimates will put it somewhere between 10 to 15% worse overall. IBM put it slightly higher at 20% worse performance using RAID 5. The performance degradation is linear too. The more drives I have in the array, the better read performance is, but the worse write performance becomes.
As you said NLO attachments maybe a good candidate for RAID 5, but only if last accessed file modification is turned off. That of couse assumes the DOAS repository is on a different volume, but a very valid point.
I am planning to roll-out DAOS on our Domino 8.5 mail server. Could you please get clarification for below points.
Question>
We have 3 mail servers in domino cluster, i.e, 2 internal user (windows based) + 1 mails Server in DMZ (linux)., In such scenario do we need to implement DAOS on each server?
If we enabled DAOS on one of internal server and the DAOS attachments moved to DAOS Directory on same machine. Accordingly attachment references link will be updated in mails of individual databases. Once this mail DB gets replicated on cluster, how does the attachment links will be accessible in cluster. Do we need to mapped the DAOS directory on clustered machine?
when mail is access from external DMZ machine, can attachments will be accessible ?
Thanks & Regards,
Darren, thanks for your helpful posts from the field.
I just want to check your RAID 5 comment and make sure I understand the context of this RAID 5 wickedness.
TXN logs certainly should be on RAID 1 or RAID 10. Domino data will also benefit from RAID 10 on a large busy server (especially if a lot of the sequential reads of file attachments have moved to DAOS, skewing the I/O more towards writes).
But DAOS NLO files themselves would seem like an ideal usage case for RAID 5. They're written once, read at occasional intervals and eventually deleted. No need for highly optimized I/O there that I can see.
Beyond that, have you attempted to optimze block/stripe sizes for these different data types? Starting in 8.5, TXN logging can auto-adjust itself to the stripe size.
How about with NLO files? It seems that setting the block size around the median point of your distribution curve for NLO file sizes might make sense. But you'd have to make some big guesses to plan that in advance, based on DAOS estimator output.
If the estimator actually produced graphical output, we'd be able to set the minimum DAOS file size and the block size for the volume much more accurately.