2 Million IOPS – It Takes The Right Architecture
In a recent blog post titled 2 Million IOPS – It Takes A Flash Village, George Crump discussed a joint project for Dell, Brocade, Emulex, and Violin Memory.
These companies have built a high-end infrastructure of compute, network, and storage components. To quote George Crump: “The environment was tuned for performance without much regard to cost.”
They claim that their crowning achievement was going over 2M IOPS of 8KB read operations.
Crump goes on to discuss the how and why of this achievement, and the reasons that one might pursue such results. One of the main points raised was that the Brocade FC switch was instrumental to the success of the project because it enabled creating a true shared environment over the four distinct Violin memory arrays.
I think someone may be looking at things backwards.
Think about it this way: if you start with the wrong storage architecture, you are automatically forced to make up for it with high end connectivity. You’re also forced to disregard cost. That way, you end up unnecessarily paying millions of dollars — we’re talking about the cost for top-of-the-line Gen5 connectivity and about 70TB of raw SLC flash.
But if you base your storage on a true scale-out architecture, you can use components already in place in every data center to get to the same and even better results.
The Kaminario K2 all-flash scale-out storage array gives you exactly that with the bonus of a much better ROI.
K2 already showed the capability of doing 2M IOPS from a single system over a year ago. And guess what? We’ve done it again.
Let’s take a look at the cost breakdown:
Storage – 1x 86TB Kaminario K2 (at $730K)
Compute – 14x Dell M610 blade servers (at $4K each)
Network – 2x Qlogic 8GB FC HBAs per server (at $0.5K each).
The Grand Total: $800K
The servers were directly attached to the Kaminario K2. Since the K2 is a single system, all of the servers shared the available capacity with no need to interconnect. Since the infrastructure is on single storage system (rather than four), TCO is automatically better. There is no need for complex tuning, which Storage Switzerland was forced to do for their project. Your expenses are significantly more optimized — you only buy what you need, and can scale when you grow. Furthermore, load balancing and application changes are not necessary when you grow; the applications are automatically spread out across the available resources when you set up your system.
This system is able to achieve 2.5 million IOPS, as shown by the IOmeter:
But what if you don’t need those 2M IOPS? No problem. K2 is a true scale-out architecture: whether you need 10TB or 100TB you get the same benefits. Only fitted for your needs and adapted for you budget! Check out some of our results in the table below.
|4KB read||2.5M IOPS|
|8KB read||2.18M IOPS|
|8KB 70% read||1.8M IOPS|
|8KB Write||1M IOPS|
|128KB Read||43 GB/Sec|
|128KB Write||28 GB/Sec|
|128KB 70% read||41 GB/Sec|