Garbage Collection Ruins SSD Performance for Virtualized Environments.
- Thread starter CougTek
- Start date
ddrueding
Fixture
I didn't read in the article what environments they were testing in (the images weren't loading)?
Mercutio
Fatwah on Western Digital
Given the explanation in the article, perhaps the best practice is to dedicate an SSD to each guest?
CougTek
Hairy Aussie
Given the price of enterprise SSD and the fact that you'll want redundancy (RAID1), the cost would be prohibitive. For instance, my two main virtualization nodes each have ~10 VM each. That would mean 10 SSDs without any redundancy or 20 with RAID1. For VMs needing more capacity than a single SSD (like a busy file server), it would also mean dedicating a RAID array to a single VM. Again, $$$.
I'll have to compare the 1 SSD per VM + replication software to a slower backup server versus running everything from a high-end SAN with mecanical disks (like a 3PAR, an EMC or others). The SSD solution will be faster, but at what cost?
Mercutio
Fatwah on Western Digital
I'm sure Enterprise SSD manufacturers would love the idea.
The biggest single production host I have runs nine guests. If I *had* to, that machine does have the capacity to run enough drives that I could feasibly run each on a dedicated drive, but in some cases the machines are virtual precisely because they don't need to be running on bare metal. Those machines get a daily full backup and if they have continuous changes (most of them don't), I run Crashplan as well.
AFAIK even RAID1 is as yet unsupported on the drive/controller level by anything outside enterprise storage systems, though obviously the rules are different if you're buying HP or IBM or EMC. At that point you're probably in for a penny or a pound anyway. EMC just got more ammunition to sell you a VMAX 20K instead of a 10K.
The biggest single production host I have runs nine guests. If I *had* to, that machine does have the capacity to run enough drives that I could feasibly run each on a dedicated drive, but in some cases the machines are virtual precisely because they don't need to be running on bare metal. Those machines get a daily full backup and if they have continuous changes (most of them don't), I run Crashplan as well.
AFAIK even RAID1 is as yet unsupported on the drive/controller level by anything outside enterprise storage systems, though obviously the rules are different if you're buying HP or IBM or EMC. At that point you're probably in for a penny or a pound anyway. EMC just got more ammunition to sell you a VMAX 20K instead of a 10K.
ddrueding
Fixture
Not sure if anyone else here is running Veeam or some other real-time-ish backup solution? I'm grabbing snapshots of production VMs every 30 minutes or so without needing too much overhead (but the 10GbE would help). I don't bother with RAID1 on the SSDs that run the VMs, and the big-ish ones (databases and such) do have their own SSD.
Mercutio
Fatwah on Western Digital
As long as my databases snapshots are happening and getting dumped to where they can be backed up, I generally don't have to care too much about the machine configuration on most of my VMs. The most obnoxious thing for me is finding a vendor puts their stuff in an oddball like, say, Gupta SQLbase, that I'm not terrible familiar with.
Howell
Storage? I am Storage!
I don't think garbage collection specifically is the issue here. Even the problems they describe are more related to wear leveling specifically; though write amplification generally would cause a problem too. At a host OS level the VMs having their blocks moved is not any different from application data files having their blocks moved due to wear leveling. The controller doesn't care that Adode premier is churning through a file when it decides that the blocks hosting your outlook ost need to be leveled out.
What might help the problem is allowing for a greater delta of wear level between pages and over-provisioning in general. The controller could keep track of which blocks had been read from recently and increase the delta if the files had been read.
What might help the problem is allowing for a greater delta of wear level between pages and over-provisioning in general. The controller could keep track of which blocks had been read from recently and increase the delta if the files had been read.