AIX

Hi Everyone,

I was wondering if anyone has any experience or information regarding mirroring volume group which has the filesystem running Oracle Database and Application E-business suite. Will this deteriorate the performance for the users ?


Secondly, if one of mirrored disk fails, will there be any impact on performance ?

thanks

Mirrors have double write overhead, a minimal issue if there are two cables and controllers involved, but with lots of I/o going on, there it is. On the flip side, your query bandwidth is doubled, as you can read either side. Of course, it is the same controllers, cables, spindles, heads, so total bandwidth for reading is not doubled, just congestion is avoided if a disk might be read in two places with the heads zipping back and forth.

Mirrors can outperform other raids because there is no parity overhead on write, and no parity overhead on a failed read.

If a volume is down, there is no double writing, which helps as there are no longer 2 places to read. When rebuilding the volume, unlike raid it is just a simple copy, but that is when the I/O capabilities are most reduced. While it is a risk, some shops hold off rebuilding volumes until off hours or the weekend.

Mirroring and RAID can happen at different levels of the software and hardware hierarchy. When at a lower level, special hardware can reduce the load. When at a higher level, it is easier to control the cost of activities like rebuilding so there is not a noticable load on the system.

Either way, one critical aspect of these redundant systems is the monitoring, so one failure is ignored until 2 cannot be ignored. Another is competence in repair work! It is tragic to pull the live mirror thinking it is the dead mirror . . . .

The vast majority of Databases that I see are hosted on SAN Disk arrays, and in that case I'd suggest that you let these arrays manage the RAID level on the disk. The only cases in those circumstances that you would mirror the disks at an AIx level would be for redundancy across disk arrays.

If ku are using local disks then you need to be careful how you lay out your volumes to avoid disks running hot. In a mirrored pair, a write is only concluded when both disks report that the write is complete, so hat can be slower than a single disk. However Reads are from any member of the mirror so can be faster.

If you can segregate high churn and query from low churn and query (work in progress versus quiescent history), RAID alone is fine for history, but mirror is better for both high churn and high query.

For instance, RAID with 5 drives 1-5 is striped so drive 5 block zero is parity for superblock 0, and drive 5 block 1 starts superblock1. Every write writes all 5 drives and every read can only be had from one device, but sequential reading is striped 5 ways. Writing is striped 4 ways since parity takes the 5th drive.

In practice, SAN mirrors are often raid at a lower level, so there can be a hidden cost, since you cannot reclaim the raid %.