November 18, 2003 at 4:16 am
Buying a new data warehouse server, looking at storage.
Option #1 is direct attach, probably a three 6402 raid controllers with 4454 split bus cabinets, so each channel has 7 drives.
Option #2 is a MSA1000 modular storage processor system, dual fiber HBA's in the CPU and a single storage processor in the MSA1000 with three shelves of disks.
In both cases the drives would be 15k 36G drives.
Goal is performance and reliability but not looking for NSPOF.
------------------
We now have a pair of Ultra 3 direct attached and we are looking for a 3-4x bump in performace. We seem to saturate now at about 500-800 io/s.
November 21, 2003 at 8:00 am
This was removed by the editor as SPAM
November 21, 2003 at 8:59 am
The MSA should give you much more than 800 i/o's a second. A 15k drive should do around 133 i/o's by it's self. One thing to look at is the raid level as well. It takes only 2 ops to write to raid 1 or raid 10 but 4 ops on raid 5. I have deployed an MSA 1000 and it performed better than direct attached system we had been using.
I don't know if running three raid controllers will put you over the top of an MSA. I've personally never had that config in house. I would also talk to your HP rep about numbers i/o's and MB/sec.
Wes
November 21, 2003 at 9:04 am
THanks. I keep pressing both HP and EMC for some apples to apples comparisons (e.g. IOMeter runs) but haven't gotten anywhere yet. They both want to quote "we can do 40,000 IO/s to cache", which is a significant number but not really a way to compare subsystems.
Anyone got some of these subsystems and want to post some IOMeter runs? I'll offer our Dell PV220 and old PV650, and newer Compaq (mumble -- have to look) raid controller direct attached?
November 21, 2003 at 10:06 am
With the same number of drives in each configuration, direct attached storage with separate controllers will perform better and cost less.
If this is a true read-only data warehouse, then you don't need to optimize the DSS database's tran log, but you should monitor tempdb usage. If tempdb gets a lot of use, then put its data on a dedicated RAID 10 or RAID 0 array and its tran log on a dedicated RAID 1.
The latest 15Krpm drives can sustain 125 random I/Os/sec, so you'll need only 20 drives to get 2500 I/Os/sec. If the database size is less than 300GB, then twenty 36GB drives in a RAID 10 array would work well. If the database size is larger, then RAID 5 would be okay if you can stand the additional time it will take to load the data. We've even used RAID 0 for TB-size data warehouses where a day's downtime to repair and reload was tolerable. Use one 6404 controller connected to two MSA30 boxes; you don't really need to split the buses with only ten drives active (plus one hot spare) in each enclosure, but you can drive four channels with the 6404 controller. Set the controller's cache to 100% read except when loading, when you'll reconfigure it to 100% write.
If you don't have enough room inside the server for the os, tempdb, and tran log arrays (i.e. you're using a DL or BL rather than an ML box), then use another MSA30 with a split backplane attached to two separate controllers for tran logs (two drives) and tempdb data (? drives). Set the cache of the tran log's controller to 100% write, and the tempdb controller to a mix of read/write.
--Jonathan
--Jonathan
Viewing 5 posts - 1 through 4 (of 4 total)
You must be logged in to reply to this topic. Login to reply