Poor SQL performance with raid10

Question

Maybe I'll try my luck here. Server Fault site was quite hostile for non-professional server user.

I recently bought HP ProLiant ML310e Gen8 v2 server with 8G's RAM and 4 HP 1TB 6G 7.2k rpm SATA (3.5in) Non-Hot Plug MDL HDD disks.

It was surprise to me that HP didn't provide debian drivers for raid controller and I had to go with software raid. I installed Debian with it's own software raid and raid level 10.

After installation I am experiencing huge performance loss and it looks like I managed to track it to MySQL operations.

Server is used mainly as Apache2 server with MySQL database.

Here is simple script that I ran to measure SQL operation times with server and my desktop PC.

$sql = "SELECT id FROM c_event";
$events = $db->getall($sql);
foreach($events as $item)
    $db->query("INSERT table1 set id = {$item['id']}");
foreach($events as $item)
    $db->query("DELETE FROM table1 WHERE id = {$item['id']}");

I compared results with my desktop PC with Intel Q6600, 4GB of ram and 120G SSD disk.

• desktop completed in 45 secs
• server in 2 hours, 46 mins and 18 secs.

Difference is totally insane.

Server raid performance seems ok to me, as disks ain't top quality ones.

hdparm -t /dev/md1
/dev/md1:
 Timing buffered disk reads: 1024 MB in  3.00 seconds = 341.04 MB/sec

I was wondering if there was something misconfigured in my server or is it totally just lack of disk performance.

For comparsion, I had possibility to run same test on ML350p Gen8 E5-2609 2.4Ghz server with 4 HP 146GB 6G SAS 15K 2.5in SC ENT HDD disks in raid 10 mode. Test took about 10 seconds.

Answer 1

Check your configuration. Do SHOW VARIABLES; on each machine, then diff the output. It smells like one of these is different: innodb_buffer_pool_size sync_binlog innodb_flush_binlog_on_trx_commit or maybe something else.

Now to be hostile... RAID-10 (or -5) without a battery-backed-write-cache is hardly worth having. Sure, you gain recovery from a disk crash (a rare event), but meanwhile, you don't get the write speedup of the write cache.