Field CTO for the Americas at Infinidat.
getty
Performance has always been a product attribute that IT vendors brag about. For decades, almost all attention was focused on CPUs — specifically how many millions of instructions per second (MIPS) they could process. Unfortunately, product specifications like MIPS don’t reflect actual customer value.
Two aspects have changed over the years that allow a realignment of priorities closer to where the value of performance in IT products actually lies:
1. When it comes to the term "IT," or information technology, what matters most is the information. The technology is merely those things that are necessary to store, manage, process, analyze, move and protect your information.
2. Specific to the technology, when considering performance differences between products, software is usually more important than hardware.
There are three independent parts of information performance to consider:
1. Throughput: This defines how much information can be processed in a period of time and is usually measured in terms like megabytes per second (MB/s) or gigabytes per second (GB/s). Throughput is important if you need to process very large data items, even if the number of items is small. Application environments such as medical research or high-definition video streaming often fit this profile.
2. I/Os per second (IOPS): This counts how many individual discrete information entities can be processed per second. IOPS is important if you have a very large number of data items, even if the size of the data items is small. Most transaction processing applications fit this profile, such as online banking and cellular phone billing.
3. Latency: This is how fast a response is returned for any request. Latency is measured simply in units of time. Depending upon where in the IT infrastructure we’re measuring it, latency can be expressed in seconds, milliseconds, microseconds or nanoseconds. Anyone who has ever been on the internet understands latency. When you click on something, how long does it take for you to get a response?
Of these three, I believe our focus should be on latency. For most applications, leading solution alternatives usually offer sufficient throughput and IOPS. It’s latency where we see the biggest opportunity to impact success.
One of the biggest consumer examples is online commerce. It's often said that if a webpage does not load in three seconds or less, most people will abandon the site. It can also cost you a customer for life — simply because of bad latency.
Financial trading is another example. Financial firms use artificial intelligence (AI) to detect market changes in real time to identify trading opportunities. If a system delays even fractions of a second, the opportunity may be lost.
When machines are reacting to each other, it’s even more critical, such as in the evolving world of the Internet of Things (IoT). Many cars today have automatic braking systems. If you aren’t paying attention and don’t hit the brakes, the car will react and brake by itself. How much latency will it tolerate before it hits the brakes? There’s not a lot of data, throughput or IOPS involved, but if the latency isn’t extremely low, things will not end well.
Many IT vendors are notorious for publishing “hero numbers” for all three performance criteria. A “hero number” is when a vendor runs an in-house lab experiment using an artificial application to attempt to make their numbers as attractive as possible. Vendors then battle for bragging rights about who has the best hero numbers.
Here’s why it doesn’t matter — the claimed hero numbers for throughput and IOPS are almost always:
• An order of magnitude higher than what most real applications require. Bragging about a car that might go 500 miles per hour isn’t important if you only need to drive 70 miles per hour.
• Conducted on configurations that were optimized just for the test and are not what real users would consider to run real applications. If you are looking to buy a small SUV, test driving a Formula One race car isn’t particularly useful.
But no matter what the claimed latency hero numbers are (probably lower than what you’ll ever see in the real world), lower latency that continually tries to approach the unachievable goal of zero latency will always add value.
So, how do today’s leading-edge products deliver superior latency over the slower ones when most vendors are building their systems from many of the same commodity components? Software — particularly by using artificial intelligence (AI).
The fastest systems use AI to try to determine what data will likely be requested by an application before the data is actually requested. It’s kind of like walking into a restaurant and the waiter already knows, with high confidence, what you are likely to order and has the chef start making your meal before you’ve actually ordered.
The best AI is a subset of AI called deep learning, where the system may place multiple bets on what it thinks will be requested next, and then it learns from which option was (or was not) chosen to influence future decisions. This would be like the waiter knowing that, in the past, you ordered either a pizza, a sandwich or a salad. With that knowledge, the waiter has the chef prepare all of them as soon as you walk in the door. By being able to deliver your meal almost immediately when you order it, the latency for waiting for your meal has been reduced, possibly drastically.
So, if an IT company starts talking about performance, push for clarity and relevance to your organization. Ask if they are willing to model the latency they can deliver for your requirements. Ask about how they use software and AI to provide lower latency that can help you. It’s one thing to promise big numbers that may or may not be needed — it’s another to be a split second too late.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
