It would be wonderful if we could purchase whatever we want. I would love to have a Tier 1 storage array next to my computer. But somehow the cost of the purchase and running it just does not make sense for my own small environment.
Equation: Amdahl's Law
So, how do we decide where to focus our efforts and funds? Here is where we can use Amdahl’s Law to try to obtain a feeling for where we should focus. In Amdahl’s Law, the speed increase (s) can be calculated if we know the fraction of the time (f) we use the faster mode and the amount of speed increase (k) while in the faster mode.
Let’s assume we have two different arrays. One array has 10K drives and the other has 15K drives. Certainly, if we upgrade the array with 10K drives we will get a 50% increase in performance, right? Right?
Well, not really. Here’s why. Let’s assume that our application processes data sequentially. For example, let’s assume we have two arrays and we shadow data between the two arrays. In that case, if we drop the member on the array with 10K drives and then use that member for backup operations, most of our backups will be fairly sequential in nature. We do a large block transfer (32K or 64K in size) and we tend to start at the beginning of the volume and head to the end of the volume.
Sure fragmentation will tend to create some randomness, but as long as the volume is not TOO fragmented, it tends to read from start and reads until the end of the volume. So the reads are very sequential in nature.
How does that change things? Well, in this situation, the storage actually spends very little time doing random I/O requests. It does not matter all that much how fast the heads spin. The heads are usually ready to read the next segment. And most arrays will prestage data into the array cache. Thus, having faster drives just does not matter. So, in this case of Amdahl’s Law shows a tiny speed increase (s) because we also spend very little time (f) in the increased mode. Though the speed increase while seeking is almost 50% greater (0.5), the fact is due to the little amount of time spent seeking, the actual improvement for performance is minuscule.
But what if the application does a huge amount of random I/O requests (such as for interactive lookups of data within a database). In that case, the improvement in performance, while not the full 50% (0.5), it is high enough to warrant the change. For example, those same two arrays, with an equal workload against them (using Host Based Volume Shadowing), will tend to see about a 20 to 25% increase against the array with 15K drives. OpenVMS automatically prefers the shadowset member with lower I/O requests and a lower latency. That random I/O request workload presented against the 15K based member will be better able to quickly respond and return data.
So, we can use Amdahl’s Law when we examine service centers throughout the computing and storage subsystem to try to identify the Best “Bang for the Buck”. By examining the potential improvement we can improve the return on investment (ROI).
And that’s the magic of Amdahl’s Law for storage performance analysis. The next blog entry will examine the impact of resource utilization on the responsiveness of work through the service centers.
With 30 years experience, John harmonizes OpenVMS, Linux and VMware with storage.
