More VMs vs. “Rust-on-a-Platter”

One of my constant mantras is “SSDs are not hard drives!” This idea might seem weird at first, but it’s become apparent to me that we have to stop thinking in terms of rust-on-a-platter storage. Storage needs are continually moving away from single-threaded, serial-access workloads to random access, multi-threaded, parallel workloads. You just can’t keep up with that using the technological equivalent of a record player in a world full of iPods.

I’ve been a user of VMWare since it was at version .9x-something. If you were an earlier adopter like me, you quickly realized “we aren’t in Kansas, anymore”. Once you get more than a couple of VMs going, things bog down, and it’s not just a matter of RAM and CPU.  You can carve out CPU cores, ration out RAM and still you are down to that hard drive being the bottle neck!

A single hard drive just can’t handle that much random traffic, and the problem is only partly dealt with by using large RAID arrays. To deal with the storage workloads of modern virtual-machine, virtual desktop and remote-application delivery systems … you need storage designed from the beginning to “do random.” Intel SSD, PCIe and NVMe technologies are the answer.

For instance, we tested comparable VMWare/VSAN configurations, and found that SSD storage clusters outshine HDD performance. This results in either being able to host a given number of VMs with better performance, or a greater number of VMs at a given level of performance. How does 300% better sound?  Keep reading for details.

Performance remains more consistent as well. That’s not surprising, since Intel SSD products are famed for their “consistently amazing” performance. Check out a pretty typical test workload against another brand/model:

1.png

Published specifications tend to focus only on absolute maximum IOPS. You can see above that the competitors drive (barely) edges out the Intel P3700 in this respect. That’s a bit like noting only the top speed of an automobile as a metric, while ignoring every other aspect of power, performance and handing. Other than the top end number, you can see that the Intel outclasses the competitor under “actual driving conditions” … greater throughput, and consistent performance while we’re at it! 

Currently supported configurations don’t (yet, stay tuned!) allow for the use of SSD drives as primary VMWare VSAN storage, but we decided to give it a go anyway. The results were amazing compared to HDD primary storage. Let’s take a look at the high points:

2.png

As we start to add VMs, we can see that the Intel SSD serves up the data faster in every case. The SSD cluster provide twice the throughput at 32 VMs than the HDD cluster.

3.png

VMs need to be responsive, so latency is a pretty big deal. In this case we can see lower latencies with the Intel SSD cluster compared to the HDD. Impressively, we even maintained sub-millisecond latencies with 8 VMs! By the time we get to 32 VMs, the HDD cluster is clearly (and dramatically) getting sluggish, while the SSD cluster is barely starting to breathe hard.


4.png


As impressive as average latency results are, it’s equally important to measure maximum latencies under load. These are the “worst case” moments which can plague virtual desktop and remote application delivery. Dang, that’s an impressive set of results, which clearly shows how the Intel SSD solution offers a consistent level of superior performance.

As a group these results show up to 200% throughput increases with SDD over HDD as VMs scale, with up to 200% lower average latency and up to 300% lower maximum latency. These factors are critical for hosts running larger numbers of VMs , hosted applications or virtual desktops.

To sum it up: SSDs are not Hard Drives, and that’s a good thing! Making this clear is one of my pet projects, so check back for more reasons why you need give up on spinning rust-on-a-platter storage.

__

Results based on internal Intel analysis and testing provided for informational purposes only. Any difference in system hardware or software design or configuration may affect actual performance. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as Iometer, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. Source: Internal Testing