Abstract
We will, in this thesis, closely investigate the performance of SSDs and
magnetic disks. Looking at physical attributes of these two disk technolo-
gies, we will try to give an understanding of how they perform in dif-
ferent scenarios based on real workloads. Based on what we learn from
investigating the performance in these scenarios, we will try to get an un-
derstanding of what performance issues the different storage technologies
faces. We will look at how optimizations in Operating System (OS) and at
application-level have an impact on SSDs and magnetic disks, as well as
what we can do to achieve best possible performance.
In our benchmarks, we show that early generation SSD suffer from write
amplification, and will, because of that, have problems achieving good
performance on random write operations. By introducing a log-based
file system, we see that we can remove the effect of write amplification
received by random write operations, and give us a more stable perfor-
mance. We discuss, and benchmark, possible alternatives for optimizing
existing applications for magnetic disks, by introducing SSDs, as an alter-
native to magnetic disks as storage device. By going from high-end mag-
netic disks to SSDs, we see that we, in some cases, can achieve orders of
magnitude better performance, without changing application level code.