We're a little behind in the virtualization game, I'll have to admit. We currently have nothing virtualized, but I'm trying to change that.
We want to get our feet wet with virtualization and get a feel for it, so I am going to propose for next year we get a server and virtualize 3 rather under-utilised (and incredibly old) non-critical servers.
I know virtualization is all about SANs, but unfortunately we don't have one (all our servers are DAS). The question I have is if I get a server (probably a HP DL380 of some description) with DAS, are the guest VMs (under-utilised don't forget) going to suffer badly as a result? My current thinking is to put Hyper-V Server 2008 R2 as the host, if it makes any difference.
I'll give a quick run down of what the 3 servers do below. I know this isn't an exact science, but roughly speaking - do you think DAS will be OK?
Server 1
Runs a legacy HR system which is accessed maybe once or twice a week. This has a SQL Server 2000 backend which also runs on this server. This thing is archaic (I don't know when HP stopped branding ProLiants Compaq, but this proudly displays the Compaq logo on the front).
Server 2
Runs some emulated printers for our AS/400 and prints the output to some real printers. This runs happily on an oldish PC with W2K3 and 2GB RAM.
Server 3
Runs a server for a BI product, however all it actually does is at 4:00AM imports a ton of data from our AS/400 and compiles it into a file (this takes about half an hour). About 15 users then periodically (maybe once every few days) download this compiled file (~100MB) to their local computer. The download is user initiated, so there's no real defined access pattern.
This is also on a PC running 2K3 but with 3GB RAM.
Update
We've looked at virtualization several times now, but because of our lack of SAN and the many advantages it brings, we have dismissed virtualization as an option each time.
I know some of the pitfalls of putting these VMs on DAS (lack of HA etc) - and that is OK for these machines. If the host were to fail, it would be an inconvenience rather than a full blown disaster.
I'm looking at this more of a consolidation project rather than anything else. The 3 servers are slated for replacement next year anyway, and like I said they are getting rather old, sit mostly idle and are big tower machines using a lot of space and electricity. In my opinion, 3 new servers (however basic) is still overkill for machines that sit doing very little all day.
Your question can be answered partly by looking at the performance counters for I/O in the machines that you're trying to migrate to virtual. Look at how many I/O operations per second (IOPS) each requires. Remember that each physical disk is good for a few tens of IOPS. Now size your virtualization system to support that level of I/O. You'll know how many "spindles" you need. Use a RAID controller to spread the load across the disks.
Note that none of this said anything about a SAN. SANs are great, but they mostly provide things that you didn't mention: high availability, backup, storage tiering, fault tolerance, cross-site disaster recovery, rapid deployment, etc.
The most notable part of all of that is that if you use DAS, the failure of the physical server that the DAS is attached to constitutes a failure of all the VMs that were hosted on it. With a SAN, using a cluster of compute hosts all attached to the SAN, you can survive the failure of the physical host by moving the VM to a new host.
Another strategy for surviving failure of that physical host is to back up the apps and OSes within the VMs frequently, or at least frequently enough that you won't mind losing things that happened after the last backup.
Your choice. DAS can give you the performance you need (given what you have described.)
1. Performance In my experience, SAN storage and DAS storage have different characteristics which make performance hard to measure objectively, since workloads tend not to rely on just one metric (e.g. access time), but rather a variable blend of access time, throughput, read vs write, etc. It's a very complex picture. The main benefits of a SAN, IMO, are having your data available to several hosts simultaneously (allowing for clustering/high availability), and unifying the storage of your data allowing for easier management and maintenance. So long story short - I wouldn't sweat performance all that much for DAS.
2. SAN Availability & Cost
You don't have to go with off-the-shelf discrete SAN solutions. Many folks are building their own since a collection of regular server parts can get you a fully functional SAN system. Here's one route to that:
Vendor support can be purchased for OpenFiler, and your hardware, if your company insists on these things. You can also configure multiple OpenFiler hosts with DRDB to give you high availability and redundancy options.
We've run one of these in production for over a year now, and it's been very solid.
DAS normally, should be faster than some FC SANs, and definitely faster than iSCSI SANs. You're getting a SAS link speed (3Gbps and in the newer boxes - 6Gbps)
The downside to DAS is not speed or stability, it's scalability, because the amount of ports on a DAS is limited, so you can't connect too many hosts to it.
When we started with virtualization in 2007 we had drbd in mind. So we started with a server-pair that had 5 internal 3,5" SAS-disks (configured as RAID 5 providing 500 GB netto storage). Since our database-servers run in a different cluster (SAN-attached) our virtual systems don't do much disk io - which makes this a possible scenario - even with 100 MBit Ethernet for drbd replication.
With that setup we were able to provide over 20 HA-VMs (XEN paravirtualized). Disk IO or disk performance was never a problem. DAS is even faster nowadays than local storage. So for your setup with just three low performance, low disk-io servers I would say that even DAS is oversized.
Since you use mainly W2K you could play with Linux KVM or Windows Hyper-V. Or Windows-Admins love the latter (and use DAS attached to two servers themselves). With Linux KVM I would use DRBD as ha-storage backend (DAS or internal disks).
Just a note on SAN/DAS: We attached some of our Virtualization Servers to SAN - for those guests that now run databases with heavy IO. But: SAN is not faster than the internal disks - it just distributes io to more disks - the same can be true for DAS.
I see the real advantage on SAN in a different area: It uses its own switching fabric, which has (at my work) nothing to do with the IP routers and switches - and is much more stable. From a cluster/ha point of view you get another - really independent - connection.