PCMark 8 was originally introduced in 2013 and the storage test uses only one thread to do everything. We are by no means living in the single-core era thanks to corporations like AMD offering CPUs with higher core counts than ever before. PCMark 10 Storage uses all the CPU cores available on the platform being tested and has been validated to support up to 5GB/s bandwidth. About half the available cores/threads are being used for generating the data needed for I/O and the other half are tasked with sending out I/Os. The I/O in both PCMark 8 and PCMark 10 is asynchronous. That means that the thread sending an I/O does not sit waiting for it to complete, but can instead queue more I/O to match the queue depth in the recorded trace. The CPU thread count used by the benchmark does not equal to the queue depth seen by the storage device. Most modern software has been written to be a muilthreaded solution where it can push IOs from multiple threads. This should play top the strengths of NVMe devices that have come out in recent years as they were designed to handle multiple queues at the same time. So, the take home message here is that PCMark 10 takes advantage of all available threads and uses newer real-world traces.
The first test that we will be running is the PCMark 10 Full System Drive Benchmark. This benchmark uses a wide-ranging set of real-world traces from popular applications and common tasks to fully test the performance of the fastest modern drives. The goal of the benchmark is to show meaningful real-world performance differences between fast storage technologies such as SATA, NVMe, and Intel’s Optane. The Full System Drive Benchmark uses 23 traces, running 3 passes with each trace. Some of the traces include booting Windows 10 and start times for games titles Battlefield V, Call of Duty Black Ops 4, and Overwatch. It typically takes an hour to run and the number of bytes written to the drive during test with default settings is 204 GB.
On the full system drive benchmark the results had the SK hynix Gold P31 SSD finishing with an overall score of 2078 points and that makes it finish in 3rd place and it beat out some very strong competitors.
The PCMark 10 Storage Benchmarks overall score is calculated from the bandwidth and average access time sub-scores. The bandwidth subtest is defined by UL in PCMark 10 as bandwidth = bytes / busy_time_for_read_and_write. Here the Gold P31 NVMe SSD averaged 333.8 MB/s on the 1TB drive and 319.9 MB/s on the 500GB drive.
The final subtest result is the average access time. During a trace playback in PCMark 10, the start and end time is measured for each I/O. So, the average access time is derived from the end time of an I/O subtracted from the start time of that operation. The SK hynix Gold P31 1TB and 500GB drive had an average access time of 81 and 85 microseconds, respectively.
The second and final test that we will be running is the PCMark 10 Quick System Drive Benchmark. The Quick System Drive Benchmark is a shorter test with a smaller set of less demanding real-world traces. This benchmark is better suited for testing entry performance level and lower capacity drives in less demanding use scenarios. The benchmark is designed to measure the performance of small system drives from traditional spinning drives (HDDs) at the low end and entry-level PCI Express SSDs at the high end. The Quick System Drive Benchmark uses 6 traces, running 3 passes with each trace. Three of the traces involve copying JPEG images and the other three use Microsoft Excel, Adobe Illustrator, and Adobe Photoshop. It typically takes 20 minutes to run and the number of bytes written to the drive during test with default settings is 23 GB.
On this lighter workload the SK hynix Gold P31 drives dropped down a bit, but remained in the upper half of of the results.
In the PCMark 10 quick storage drive benchmark bandwidth test the P31 1TB came in at 268 MB/s and the P31 500GB finished at 260 MB/s.
The average access time on the Gold P31 1TB SSD was 60 microseconds and the Gold P31 500GB was just a touch slower than that at 62 microseconds on the quick system drive benchmark workload.