The decision to move to virtualization-using-KVM as our standard way of deploying servers was really a success, given the cost savings for the past 2 years. The only downside is the performance hit in intensive disk IO workloads.
Some disk IO issues were already addressed in the application side (e.g. use cache, tmpfs, etc., smaller logs) but it’s apparent that if we want our deployment to be more “denser”, we have to find alternatives for our current storage back-end. Probably not a total replacement but more of a hybrid approach.
Solid State Drives is probably the the best option. It is cheaper compared against Storage Area Networks. I like the idea even more because it’s a simple drop-in replacement to our current SAS/SATA drives compared against maintaining additional hardware. Besides, my team does not have the luxury of “unlimited” budgets.
After a lengthy discussion with my MD, he approved to perform some tests first to see if SSD route is feasible for us. I chose to use 4 120GB Intel SSD 320s. The plan was to setup these 4 drives in a RAID 10 array and see if how how many virtual machines it can handle.
I chose Intel because it’s SSDs are more reliable among the brands in the market today. If performance is the primary requirement, I’d choose a SSD with a SandForce controller (maybe OCZ) but its not, its reliability.
The plan was to set-up a RAID 10 array of four 320s. But since our supplier can only provide us with 3 drives at the time we ordered, I decided to go with a RAID 0 array of 2 drives instead. I can’t wait for the 4th drive. (It turned out to be a good decision because the 4th drive arrived after 2 months!).
The Intel 320s write endurance, 160GB version, are rated at 15TB. My premise was, if we’re going to write 10GB of data per day, it will take almost 5 year to reach that limit. And in theory, if it’s configured in a striped RAID array, it will be a lot longer than 5 years.
It’s been over a month since I set-up the ganeti node with the SSD storage, so I decided to check and see its total writes.
The ganeti node has been running for 45 days. /dev/sda3 is the LVM volume configured for ganeti to use. The total blocks written is 5,811,473,792 at the rate of 1,468.85 blocks per second. Since 1 block = 512 bytes, this translates to 2,975,474,581,504 bytes (2.9TB) at the rate of 752,051.2 bytes per second (752kB/s). The write rate translates to 64,977,223,680 bytes (64.5GB) of total writes per day! Uh oh…
64.5GB/day is remotely near from my premise of 10GB/day. At this rate, my RAID array will die in less than 2 years!
Uh oh indeed…
It turned out that 2 of the KVM instances that I assigned to this ganeti node are DB servers. We migrated it here a few weeks back to fix a high IO problem. A move that cost the Intel 320s a big percentage of its lifespan.
It seems that 64GB/per day is huge but apparently, it’s typical on our production servers. Here’s an iostat of one of our web servers:
I’m definitely NOT going to move this server to a SSD array anytime soon.
As a whole, the test ganeti node has been very helpful. I learned a few things that will be a big factor on what hardware we’re going to purchase.
Some points that my team must keep in mind if we’ll pursue the SSD route:
- IO workload profiling is a must (must monitor this regularly as well)
- leave write intensive VMs in HDD arrays or
- consider Intel SSD 710 ??? (high write endurance = hefty price tag)
I didn’t leave our SSD array to die that fast of course. I migrated the DB servers to a different ganeti node and replaced it with some application servers.
It decreased the writes to 672.31 blocks/sec (344kB/s), more than half of its previous rate.
Eventually, the RAID array will die of course. For how long exactly, I don’t know, > 2 years? 🙂