General rules DO have exceptions .. If you are 100% sure you won't need more RAID space or more mirror space in the future, using different HDA for different symdevs will give benefits. But as soon as you want 1 more RAID symdev (or 1 more mirror symdev) you have either to upgrade capacity or mix protection on the same HDA .. And if you choose to upgrade capacity, remember that all new symdevs you build on new capacity will be "bound" to the new drives.
"Once upon a time" you had HDA for your BCVs, HDA for your mirrored volumes, HDA for RAID-S devices .. But you create partitions in your backend. And partitioning HDAs (thus creating too many DGs) may give unpleasant results (since you aren't spreading workload on all HDAs but only on a given subset).
Statistically it's better to spread workload on a lot of HDA instead of having a perfectly working DMSP. But if you feel DMSP is mandatory for your environment (and only you can know if DMSP is really that important for your env), ask EMC to create 2 disk groups and go create mirrored devices on DG1 and raid5 devices on DG2
Thank you for your reply Stefano, It is much appreciated.
But I am still a bit confused. Even setting DMSP aside, wouldn't the improved write throughput that is inherent with RAID1 be compromised by the head movement to access RAID 5 tracks on the same physical drive.
Let's say for example that I have 820 usable 146GB drives. I want to use half of the space as RAID1 and half as RAID5. I am trying to understand why there would be a recommendation to spread RAID1 and RAID5 across all 830 drives instead of using 410 drives for RAID1 and 410 drives for RAID5. Can you help me with what might be the benefit of splitting all of the drives?
DMSP works only for mirrored devices .. Obviously having a mix of RAID and Mirror devices on the same volume may reduce benefits of DMSP. RAID volumes will cause unwanted head movement. However note that DMSP is one of the benefits offered by Mirroring .. You partially loose DMSP but keep everything else
I am in the same boat and chose to include both RAID1 and RAID5 devices on the same spindles for 1 reason. We plan on putting all of the low-IO requirement stuff onto RAID5 and high-IO on RAID1. Since each spindle can only handle so much IO, if you use it all for high-IO applications, at some point you probably won't be able to get any more performance out of the spindles even though the capacity is far from being used. But, by adding some of the applications that don't need high IO, but need lots of space and the availability of the enterprise array, we can use up both our capacity and our available performance somewhat evenly.
The other thing is that this allows you to spread your RAID1 across all spindles, really using the available drives. I hate to see 20% of my spindles screaming with capacity still available and 80% loafing along, but completely full. By distributing it all (and using SymOptimizer), I'm hoping to see the load fairly equally distributed across all spindles.
That does give me something to ponder. I see your point, if half my physical drives are only RAID1, I will probably run into myself occasionally creating my own hot spots.
But by inviting head movement to RAID5 areas on the same disk, I would be preplanning to get no better than average write throughput.
It seems like with hypers and meta-volumes it would be nearly impossible to get optimum throughput from these drives in a pure RAID1 configuration.
But I am still a bit confused. Even setting DMSP aside, wouldn't the improved write throughput that is inherent with RAID1 be compromised by the head movement to access RAID 5 tracks on the same physical drive.
Let's talk about write performances and RAID devices ....
Usually you write in cache, thus write performances of the backend isn't usually an issue. It may become an issue if you hit WP limit. But again raid5 MAY perform better then mirrored devices while writing since when you write on a mirrored device you write the same data on two devices, while when you write on RAID devices you spread the workload on at least 4 drives. Thus if your application isn't that write intensive (up to 30%) RAID5 3+1 may outperform mirrored devices while testing raw performances. However I do agree that in real world examples, Mirrored devices outperforms RAID5 almost ever.
Yeah, with shared resources it is all about trade offs and compromises. You can go for maximum performance, but have a risk of hot spots and/or not being able to fully utilize your capacity. Or, you can go for reduced performance, but more utilization of capacity. The other thing is manageability. You can get some very good performance and use the most of your capacity with some manageability issues. By this I mean going with small RAID1 hypers distributed across all spindles, creating metas to present to the host and then do host based mirroring. Then, even if a host is really hammering a particular area on a filesystem or device, that IO is spread so thinly across all spindles that you reduce the risk of hot spots and have the best chance of using your capacity. But, on the flip side, you probably don't make the best use of the DMX cache and built in intelligence and managing stripes and such can be a pain.
This is also based on the premise that 2 layers of striping, array and host, is acceptable. I know that's a somewhat hotly debated topic.
xe2sdc
4 Operator
•
2.8K Posts
0
December 2nd, 2008 01:00
"Once upon a time" you had HDA for your BCVs, HDA for your mirrored volumes, HDA for RAID-S devices .. But you create partitions in your backend. And partitioning HDAs (thus creating too many DGs) may give unpleasant results (since you aren't spreading workload on all HDAs but only on a given subset).
Statistically it's better to spread workload on a lot of HDA instead of having a perfectly working DMSP. But if you feel DMSP is mandatory for your environment (and only you can know if DMSP is really that important for your env), ask EMC to create 2 disk groups and go create mirrored devices on DG1 and raid5 devices on DG2
mossmand
4 Posts
0
December 1st, 2008 08:00
But I am still a bit confused. Even setting DMSP aside, wouldn't the improved write throughput that is inherent with RAID1 be compromised by the head movement to access RAID 5 tracks on the same physical drive.
Let's say for example that I have 820 usable 146GB drives. I want to use half of the space as RAID1 and half as RAID5. I am trying to understand why there would be a recommendation to spread RAID1 and RAID5 across all 830 drives instead of using 410 drives for RAID1 and 410 drives for RAID5. Can you help me with what might be the benefit of splitting all of the drives?
xe2sdc
4 Operator
•
2.8K Posts
0
December 1st, 2008 08:00
Trouphaz
1 Rookie
•
70 Posts
1
December 1st, 2008 13:00
The other thing is that this allows you to spread your RAID1 across all spindles, really using the available drives. I hate to see 20% of my spindles screaming with capacity still available and 80% loafing along, but completely full. By distributing it all (and using SymOptimizer), I'm hoping to see the load fairly equally distributed across all spindles.
mossmand
4 Posts
0
December 1st, 2008 14:00
That does give me something to ponder. I see your point, if half my physical drives are only RAID1, I will probably run into myself occasionally creating my own hot spots.
But by inviting head movement to RAID5 areas on the same disk, I would be preplanning to get no better than average write throughput.
It seems like with hypers and meta-volumes it would be nearly impossible to get optimum throughput from these drives in a pure RAID1 configuration.
I guess we just have to live with average.
Thank you for your insight.
xe2sdc
4 Operator
•
2.8K Posts
0
December 2nd, 2008 02:00
aside, wouldn't the improved write throughput that is
inherent with RAID1 be compromised by the head
movement to access RAID 5 tracks on the same physical
drive.
Let's talk about write performances and RAID devices ....
Usually you write in cache, thus write performances of the backend isn't usually an issue. It may become an issue if you hit WP limit. But again raid5 MAY perform better then mirrored devices while writing since when you write on a mirrored device you write the same data on two devices, while when you write on RAID devices you spread the workload on at least 4 drives. Thus if your application isn't that write intensive (up to 30%) RAID5 3+1 may outperform mirrored devices while testing raw performances. However I do agree that in real world examples, Mirrored devices outperforms RAID5 almost ever.
Trouphaz
1 Rookie
•
70 Posts
0
December 3rd, 2008 07:00
This is also based on the premise that 2 layers of striping, array and host, is acceptable. I know that's a somewhat hotly debated topic.