This technical brief discusses how we used standardized OLTP performance metrics and a data set that exceeds available system memory (to test storage system I/O) to compare new orders per minute (NOPM), database average response time and response time consistency on a SQL Server using SSD versus legacy HDD storage.
We used the same base hardware (server, CPUs and DRAM) with all three storage configurations:
- Legacy: 16x 300GB 15K RPM HDD configured RAID 10 (baseline configuration for comparison)
- NVMe 1: 2x 3.8TBin Mirrored Storage Spaces
- NVMe 2:4x 3.8TB in Mirrored Storage Spaces
- Complete more transactions with NVMe — more orders, more fulfillment, more to your bottom line
- Data can’t wait — SSDs with NVMe release your data’s potential like never before
- Compared to the baseline 16x HDD configuration:
- 2x NVMe SSDs supported 216X more New Orders per Minute
- 4x NVMe SSDs supported 276X more New Orders per Minute
- Databases hosted on NVMe response times were 77% to 90% lower and much more consistent
More Orders per Minute Brings More ValueSSDs are a mainstay of high-performance, low-latency IT systems. High-capacity, ultra-performance NVMe SSDs drive those systems farther and faster, processing more data and bringing more value.
More and more OLTP platforms are moving to NVMe, and the differences between the capabilities of NVMe today and what we used to think of as a performance HDD configuration are greater than ever, with the legacy configuration’s 15K RPM HDDs being painfully slow in camparison. In OLTP systems, additional transactions can represent more orders, more fulfillment or more detailed analysis — all bringing more value.
The magnitude of the difference between current NVMe designs and legacy standards is evident in Figure 1, which shows each configuration’s relative NOPM at a system load just before the test reached a stop condition. (See the How We Tested section for stop condition details.)
The NVMe configurations’ NOPM are extremely high. The 2x NVMe achieved 216X more NOPM than the baseline and the 4x NVMe achieved 276X more NOPM than the baseline.
Figure 1: Relative NOPM
NVMe Brings Faster, More Consistent ResponsesMany applications require high NOPM while quick, consistent database response (low latency) may be more important for applications that are very time-sensitive.
As shown in Figures 2a and 2b below, we calculated and compared the mean latency and the 99.9th percentile latency (a good indicator of latency consistency) at system load just before the test reached a stop condition (see the How We Tested section for stop condition details) for the three storage configurations. We used the same metrics, database and test conditions for each configuration.
Figure 2a: Mean transaction time
Figure 2b: 99.9% transaction time
Figure 2b shows that the NVMe configuation response times were more consistent than the baseline configuration (lower 99.9% transaction response time).
The two comparisons indicate that both NVMe configurations respond more quickly and more consistently than the baseline configuration. The legacy configuration shows the opposite — much higher mean transaction response time that is far less consistent.
The Bottom LineMission-critical data can’t wait. Access delays or inconsistency can be extremely costly. Using NVMe SSDs enables fast transaction processing and fast, consistent response times.
In our testing, NVMe SSDs demonstrated tremendous benefits and new capabilities for one of the most popular database management systems and most challenging workloads — Microsoft SQL Server and OLTP. Supporting immense loading, far greater NOPM with lower and more consistent latency means more orders and more transactions completed faster and more consistently.
Learn more about NVMe SSDs and their transformative effect on your business at micron.com.
How We TestedTo ensure a fair assessment of the expected maximum NOPM of each configuration, we took a configuration-specific approach. We measured each configuration’s NOPM at the maximum load the platform could reasonably support, as opposed to comparing NOPM and latency at an arbitrary load.
Prior to testing, we established stop conditions (Tables 1 and 2). As we tested, we increased the load until the test a reached a stop condition after which we stopped increasing the load and used the NOPM and latency values recorded when we reached the stop condition.
Table 1: Stop conditions1,2
Table 2: Threshold limits
- We set the stop condition for CPU utilization at 80%. Many IT organizations plan for a platform upgrade when CPU utilization reaches 50% and implement that plan when it reaches 80%.
- We sized the data set to ensure it was large enough to ensure storage I/O (data set size about 2X the memory size) but did not occupy more than 80% storage capacity.
Determining Maximum Load by ConfigurationThis section shows the test condition(s) that established each configuration’s maximum load.
Legacy Configuration Stop Condition: Average Log Write Latency
Figure 3 shows the legacy configuration’s average log disk (partition) write latency by load. The stop condition is shown in red (at which point the average log disk [partition] write latency significantly exceeds our 5ms stop condition).
Figure 3: Legacy configuration stop condition
Figure 4 shows the NOPM (vertical bars) and 99.9% response time (line) by load for the 2x NVMe configuration. There is a NOPM plateau just before the far right of the loading range (the loading value when NOPM drops is shown in red). Note too there is a significant jump in 99.9% response time at the same loading point.
Figure 4: 2x NVMe stop condition
Figure 5 shows the %CPU utilization versus loading for the 4x NVMe configuration. %CPU utilization reaches 80% on the right side of Figure 5. We used a loading point where %CPU utilization exceeds 80% and the loading level aligns to two logical users per thread.
Figure 5: 4x NVMe stop condition
Table 3: Hardware configuration
To download a pdf version of this Technical Brief, click here.