This may seem like an extraneous section to include, but I have seen quite a few people running virtual machines in non-conducive hardware setups and then wrongly thinking virtualization was not up to snuff. The obvious realities are that the hardware must be sufficient to handle one or more database servers, their operating systems, plus the host operating system to boot. So it should come as no surprise that I recommend as much CPU and memory as budget will allow. That is good advice if you are ordering hardware, but what about reusing existing hardware assets or laptops/notebooks for demos?

When trying to utilize existing hardware, just remember that if the server was not good enough for application X, then it probably will not be sufficient for X + Y or even Y + Z (unless both are much smaller than X). Basically it boils down to one simple rule:  if a server was not sufficient for a stand-alone situation, then reusing it as a virtualized server is most likely an exercise in futility. Hardware is so inexpensive these days that you should buy what you are going to need. Remember, person hours to tune a system will very often cost more than a server these days, so overspending on hardware is a good way to hedge your bets (i.e. investments).

Possibly the most egregious situation I encounter is people trying to improve sub-standard virtualization hardware via system component augmentations that are in and of themselves neither a true nor good performance option. For example, many people will try to run VMware on their notebooks for an Oracle database contained on a USB disk drive. The USB interface has 2.5-3.0 times slower throughput than SCSI or IDE channels are capable of, so this is a very poor choice. You would be much better served to simply buy a replacement disk for your laptop/notebook.

For example, upgrading my notebook from 100GB to 200GB cost $150 and two hours to move my image from source to target drive. Even if I had bought the imaging software for just that one use, the cost would have been just another $50, for a total of $200. Since even the cheapest people charge at least $50 per hour for their time, the true upgrade cost was approximately $300 ($150 for new disk drive, $50 for imaging software, and $100 for two hours time). I can nearly guarantee that this alone will improve performance by at least a factor of two. Whereas, I cannot say with any certainty that six hours tuning (i.e. $300 @ $50 per hour) will yield similarly positive results.

Another pandemic problem I encounter is someone trying to use cheap IDE disk drives on virtualized servers. Even the fastest SATA-II IDE drives are not as fast as SCSI or SAS (Serial Attached SCSI). SCSI drives spin much faster, often from 10-15,000 RPM versus just 5,400 to 7,200 for IDE (note that although Western Digital makes Raptor IDE drives which spin at 10,000 RPM, they are both expensive and relatively uncommon). So this rotational speed difference alone can translate into as much as a 100% improvement.

Furthermore, IDE disk drives have another huge disadvantage they primarily perform I/O operations serially (i.e. complete I/O operation #1 before doing I/O operation #2). Even SATA IDE drives that support Native Command Queuing (NCQ) or Tagged Command Queuing (TCQ), which permit the disk to reorder I/O requests to minimize head movement, still perform their I/O operations serially. Whereas, SCSI drives possess a processor to minimize host CPU usage and permit higher throughput of concurrent I/O requests. Also, since a virtualized server is going to need all of its CPU resources and need to be able to handle lots of concurrent I/Os from multiple operating systems, SCSI based technology is clearly the best choice.

Regardless of whether you are ordering a new virtual server or simply augmenting an existing boxs I/O bandwidth to support running multiple virtual machines and their databases, you need to choose an I/O channel that can support such high I/O loads across multiple shared resources. And as I said in the last paragraph, IDE is not a good answer SCSI reigns supreme. But what kind of SCSI based technology should we implement? Again, much of that depends upon your hardware budget. Assuming you have the minimum budget required for a proper I/O subsystem, then some good choices include:

For Direct Attached Storage (DAS) with multiple controllers

• SCSI-320

• SAS (Serial Attached SCSI)

• SCSI-640

Storage Area Networks (SAN) with multiple HBAs (host bus adapters), directors/channels, & redundant pathways

• Fibre Channel 1GFC

• iSCSI via 1Gb Ethernet

• Fibre Channel 2GFC

• Fibre Channel 4GFC

• iSCSI via 10Gb Ethernet

The entire universe of I/O subsystem choices are shown in the table on the following page with all the above stated good choices highlighted.

Therefore, it is very important that you have the proper virtualization hardware and are financially able to create an I/O subsystem to fit your needs.  To use anything less is to create more headaches and aggravation that could be easily avoided with the right equipment.

 
This is an excerpt from Oracle on VMWare: Expert tips for database virtualization by Rampant TechPress.