Benchmarking VMware Virtual HDD Performance

I've been curious about this for a while, so I decided to measure the performance of VMware virtual HDDs.

For simplicity, I used CrystalDiskMark to obtain the benchmarks.

While this software is easy to use for benchmarking, some say its accuracy isn't very high, so please take these results as a general reference.

Prerequisites

Physical Machine

• CPU: Intel Core i7 4770
• Memory: 32G
• SSD: SanDisk Ultra II 960G (Serial ATA 3.0 connection)

Virtual Machine

• CPU: 2 cores * 3 threads
• Memory: 16G
• VMware 12 Pro

Benchmark Results

Physical Machine

Before showing the benchmarks for the virtual machine, I'll provide the benchmark for accessing the SSD directly from the physical machine as a reference value.

You can see just how much "slower" it becomes.

Sequential access is clearly being bottlenecked by the Serial ATA 3.0 connection.

Virtual Machine

The following four cases were evaluated:

• Physical drive assignment - no partition
• Physical drive assignment - with partition
• Virtual drive - pre-allocated
• Virtual drive - dynamically allocated

Let's take a look at the numbers.

Here is the graph visualizing these results:

While we can see some measurement errors typical of CrystalDiskMark, the trends are clearly visible.

Discussion

Based on the measurement data above, we can conclude the following:

1. Sequential access is bottlenecked by the bus, so there is no major difference regardless of the method.
2. Usage from a physical machine is particularly fast with queued states and random writes (which you can feel during use).
3. Usage from a virtual machine with physical disk assignment is not as fast as expected, showing no major difference compared to pre-allocated virtual disks.
4. Dynamically allocated virtual disks tend to be slower compared to others.
5. Even with pre-allocated virtual disks, disk expansion is possible.
6. It seems that physical disk assignment does not offer enough advantages to compensate for the benefits of virtual disks (such as flexible expansion and portability) (this is just my personal opinion).

These were the results.

Perhaps it was a bit unexpected?

Next time, I plan to measure the performance of a virtual environment on a Mac.