Pliops makes a storage accelerator card, called the Pliops Storage Processor, aimed at cloud service providers and the very top end of enterprises that are operating at substantial scale.
Pliops claims over 20 Fortune 500 cloud and enterprise companies as customers, reflecting their focus on a problem that is most keenly felt by this end of the market.
In this market people are concerned with both performance and efficiency because at this scale a few extra percent of efficiency translates directly into performance. At smaller scale, you can quite happily deal with being less efficient to chase better performance because the extra overheads don’t matter quite so much, relatively speaking.
If you only have a few servers, wasting a bit of extra money/space/power or whatever doesn’t hurt you so much provided the value gain of the increased performance is substantial enough. It’s just a lot easier to, for example, spend an extra $200k on some new high-end servers rather than spending a year of developer time tuning code loops. The overhead as a percentage of value is relatively small.
But if you have a few thousand servers, suddenly the payoff looks a lot different. It’s worth spending some time tuning your code to avoid—or at least delay—replacing $20 million in servers. The overhead/payoff calculation starts to look a bit different, and suggests a different approach is required.
However, after this point—and this is where a lot of the top-end enterprises find themselves—companies have both invested in tuning code and in buying better hardware. There’s not a lot of room left to squeeze performance and efficiency out of what they’re already doing, so they start looking for new ideas. Not just “what we’re doing now, only more so.” Ideas for things they’re not currently doing. They want something new, something different.
Sometimes that becomes new software techniques. And sometimes it becomes new hardware techniques, like Pliops. For this market, making the existing investments in both high-end hardware and software pay off at scale is the name of the game. Adding a storage accelerator card into your existing setup is one way to achieve these sorts of gains, and that’s what Pliops is claiming they deliver for their customers. 50% cost savings, 11x more throughput and 7x less latency on MariaDB TPC-C performance, etc.
Now these kinds of approaches often filter down to the rest of the market as costs decrease in a classic experience curve fashion. We’ve seen that before with things like GPUs, flash memory, faster RAM, more efficient power supplies, and a range of software techniques like MapReduce, eventual consistency, graph databases, and so on.
But not always. We’ve seen accelerator cards before—I’m reminded of the SimpliVity approach to hyper-converged infrastructure—and sometimes they get some adoption in a particular niche but don’t really expand outside of that niche. For others, the market simply moves on so that what was once useful—say, PernixData’s memory virtualisation software—is no longer really necessary.
It’s not clear to me yet where Pliops sits. There is a general industry shift towards more specialised compute functions connecting with data directly, rather than all compute being mediated via the CPU, so Pliops could break out of its niche if it can chart a pathway into this wider world. A focus on genuine value for problems faced today is also a great way to build a footprint from which to expand.
But there’s nothing wrong with dominating a juicy niche, either, and there’s certainly plenty of money sloshing around in the hyperscale cloud and high-end of the enterprise market.
I look forward to seeing which path Pliops chooses.