Infinite I/O

Notes from our storage lab on write performance (Part II)

Posted by Sheryl Koenigsberg on Jan 30, 2017

In our last post, we looked at how Infinio's offload capability of significantly reducing the read I/O load on storage can improve write performance.   

Now that you understand the power of offload with a server-side cache, let's see what kinds of improvement we can see from more of these HCIBench tests. On the same all-flash server SAN using HCIBench, but this time at a 4k block size, you can see the improvement to a 50/50, 70/30, and 100% read workload.  


blog image IOPS improvement.png

The results from the 50/50 read/write test are similar to what we saw in our last tests: just over 50% IOPS.  A 70/30 workload (HCIBench's OLTP benchmark) shows nearly double the IOPS, both reads and writes.

The benefit to writes is clearer when we break out the read I/O from the write I/O.  Below are the results from the same test, showing the substantial improvements not just to reads (as expected) but also to writes. 

blog image IOPS read-write bracket3.png

 

Finally, let's see what the impact is to response time.  They are similarly impressive: here are the improvements to latency from the same tests.

 blog image latency improvement.png

Again, the 100% read workload shows the biggest improvement: 70%.  But the mixed workloads see substantial benefit as well: the 50/50 workload with Infinio is 40% faster and the 70/30 workload is 55% faster.  And this is while also increasing IOPS as shown above.

To summarize, for a 50/50 workload, we increase IOPS by over 50% while decreasing latency by 40%.  For a 70/30 workload, we double IOPS and reduce latency by 55%. 

And remember, there is a significant benefit to adding a read-only cache to your environment rather than a read-write cache, if your I/O patterns support it: there's no danger of losing data.  With a read-only cache, there's never any unique data in cache, so there's no need for complex and space-inefficient replication and data protection schemes.

If you'd like to see more results from our performance labs, let us know.

Think you might see improvement in your environment?

Request a trial

 

Topics: Talking Tech