Monday, May 27
Everything I Needed To Know...
About configuring cost-effective servers I learned min-maxing characters in AD&D.
Comments are disabled.
Post is locked.
About configuring cost-effective servers I learned min-maxing characters in AD&D.
Currently looking at this:
Component | Price |
Supermicro SC116TQ-R700CB 1U case | $546 |
Supermicro X9SRH-7TF motherboard | $479 |
Intel Xeon E5-1620 | $329 |
16GB DDR3 Reg. ECC x 4 | $512 |
Seagate Constellation.2 1TB x 4 | $848 |
Intel SSD 520 240GB x 4 | $1,012 |
WD Green Mobile 2TB x 2 | $338 |
$4,064 |
The motherboard includes a RAID controller, but only levels 0, 1, and 10; no 5 or 6. On the other hand, adding a decent RAID-5/6 controller costs more and delivers less than bumping up the size of the SSDs and adding a couple of low-end drives for backup so that the main array doesn't get filled with junk.
This is a similar but more pedestrian config with a dedicated RAID controller:
Component | Price |
Acme RS108TF 1U Xeon | $739 |
Intel Xeon E5-1620 | $329 |
16GB DDR3 Reg. ECC x 4 | $512 |
Seagate Constellation.2 1TB x 4 | $848 |
Intel SSD 520 180GB x 4 | $780 |
LSI SAS 9260-8i | $525 |
LSI Battery LSI00264 | $159 |
Supermicro slim DVD-ROM | $49 |
$3,941 |
The stick-in-the-mud config comes in $123 cheaper and has an extra terabyte on the main array and an extra 60GB of SSD. It's RAID-5 instead of RAID-10, but with the databases on SSD, that makes little to no difference. On the other hand, the version that takes advantage of the built-in controller also has built-in 10Gb ethernet, and has 4TB of near-line storage for backups and archives. And that means I won't need a separate backup/archive server, which was set to cost ~$2500 by itself.
I'm hoping to get a couple of these as servers, and a couple of similar boxes set up as development workstations. Just waiting for the money tree to flower...
Update: It will be a little longer before the money tree blooms, but we're deploying a similar server at my day job for stress-testing software, so I'll have a chance to evaluate the hardware soon anyway. This is one advantage of working with big data - a small test server that amounts almost to petty cash is equivalent to a very nice server for my own projects.
Posted by: Pixy Misa at
02:30 AM
| Comments (3)
| Add Comment
| Trackbacks (Suck)
Post contains 358 words, total size 5 kb.
1
As someone who works with transactional database, I'm horrified when I see mention of RAID 5.
Posted by: RickC at Monday, May 27 2013 01:40 PM (WQ6Vb)
2
We run RAID-5 and RAID-50 SSD arrays, and they work quite nicely with our big MySQL databases using InnoDB and TokuDB. We also run MongoDB and ElasticSearch on RAID-5 SSD.
A small RAID-5 array of good consumer drives like the Intel 520 series delivers worst-case random write performance equal to a 24-drive RAID-10 array of 15k RPM traditional drives, while being much cheaper, much more compact, much lower power, and delivering 10-20x the read performance. And using server-grade drives like the Intel S3700 brings up the worst case performance to roughly match the best case performance.
In the bad old days, when server memory was measured in megabytes rather than gigabytes and access times in tens of milliseconds rather than tens of microseconds, RAID-5 was database poison. These days the bottleneck is almost certainly elsewhere.
If you're talking about big monolithic databases - single Oracle instances of many terabytes, for example - then RAID-50 or even RAID-10 might still be the way to go; the costs typically associated with such systems far outweigh a few extra SSDs, even nice ones like the S3700.
For a relatively small system like mee.nu, RAID-5 is fine, but RAID-10 works out cheaper in thise case. And for truly huge systems like we run at my day job (>2PB of data), RAID-10 isn't practical. Or necessary, really; we deploy dozens of drives (SSD or HDD) at a time, so our aggregate I/O throughput across the cluster is just phenomenal. (We just deployed 204 HDDs - 712TB - for a single project.)
A small RAID-5 array of good consumer drives like the Intel 520 series delivers worst-case random write performance equal to a 24-drive RAID-10 array of 15k RPM traditional drives, while being much cheaper, much more compact, much lower power, and delivering 10-20x the read performance. And using server-grade drives like the Intel S3700 brings up the worst case performance to roughly match the best case performance.
In the bad old days, when server memory was measured in megabytes rather than gigabytes and access times in tens of milliseconds rather than tens of microseconds, RAID-5 was database poison. These days the bottleneck is almost certainly elsewhere.
If you're talking about big monolithic databases - single Oracle instances of many terabytes, for example - then RAID-50 or even RAID-10 might still be the way to go; the costs typically associated with such systems far outweigh a few extra SSDs, even nice ones like the S3700.
For a relatively small system like mee.nu, RAID-5 is fine, but RAID-10 works out cheaper in thise case. And for truly huge systems like we run at my day job (>2PB of data), RAID-10 isn't practical. Or necessary, really; we deploy dozens of drives (SSD or HDD) at a time, so our aggregate I/O throughput across the cluster is just phenomenal. (We just deployed 204 HDDs - 712TB - for a single project.)
Posted by: Pixy Misa at Tuesday, May 28 2013 03:01 PM (PiXy!)
3
Yes, I don't work on petabyte scale, but when I moved a database off a 20+-spindle SAN to a raid-1 pair and nobody noticed, that doesn't do anything to dissuade me from avoiding raid 5.
Also, I assume your day job has people who are smart enough to have UPSes so you don't have to worry about data loss due to power outages, unlike many of the people I've ever dealt with in the past. From an actual disaster recovery conversation: "We can't just buy another disk controller for this machine at Best Buy, you know." I bet you can guess that was directed at the guy who cancelled the maintenance contract with IBM "because we didn't use it last year."
Also, I assume your day job has people who are smart enough to have UPSes so you don't have to worry about data loss due to power outages, unlike many of the people I've ever dealt with in the past. From an actual disaster recovery conversation: "We can't just buy another disk controller for this machine at Best Buy, you know." I bet you can guess that was directed at the guy who cancelled the maintenance contract with IBM "because we didn't use it last year."
Posted by: RickC at Wednesday, May 29 2013 03:41 AM (WQ6Vb)
51kb generated in CPU 0.0506, elapsed 0.1451 seconds.
56 queries taking 0.1292 seconds, 346 records returned.
Powered by Minx 1.1.6c-pink.
56 queries taking 0.1292 seconds, 346 records returned.
Powered by Minx 1.1.6c-pink.