You asked for real-life examples of our corporate SSDs and professional tests. We provide you with a detailed overview of our SSD drives from our partner Truesystems. Truesystems experts built a real server and emulated absolutely real tasks that all enterprise-class SSDs face. Let's take a look at what they did!
2019 Kingston Model Lineup
Let's start with some dry theory. All Kingston SSDs can be divided into four large groups. This division is conditional, since the same drives fall into several families at once.
- SATA SSD in 2,5″, M.2 and mSATA form factors Kingston UV500 and two models of drives with NVMe interface - Kingston A1000 and Kingston KC2000;
- . The same models as in the previous group and, in addition, SATA SSD Kingston A400;
- : UV500 and KC2000;
- . Drives of the DC500 series, which became the hero of this review. The DC500 line is divided into DC500R (primary reading, 0,5 DWPD) and DC500M (mixed loads, 1,3 DWPD).
Truesystems tested with a 500GB Kingston DC960R and a 500GB Kingston DC1920M. Let's brush up on their specs:
Kingston DC500R
- Volume: 480, 960, 1920, 3840 GB
- Form factor: 2,5″, height 7mm
- Interface: SATA 3.0, 6Gb/s
- Declared performance (for 960 GB model)
- Sequential access: read - 555 MB / s, write - 525 MB / s
- Random access (4 KB block): read - 98 IOPS, write - 000 IOPS
- Latency QoS (block 4 KB, QD=1, percentile 99,9%): read - 500 µs, write - 2 ms
- Emulated sector size: 512 bytes (logical/physical)
- Resource: 0,5 DWPD
- Warranty period: 5 years
Kingston DC500M
- Volume: 480, 960, 1920, 3840 GB
- Form factor: 2,5″, height 7mm
- Interface: SATA 3.0, 6Gb/s
- Declared performance (for 1920 GB model)
- Sequential access: read - 555 MB / s, write - 520 MB / s
- Random access (4 KB block): read - 98 IOPS, write - 000 IOPS
- Latency QoS (block 4 KB, QD=1, percentile 99,9%): read - 500 µs, write - 2 ms
- Emulated sector size: 512 bytes (logical/physical)
- Resource: 1,3 DWPD
- Warranty period: 5 years
Truesystems experts noticed that Kingston drives list QoS values of total latency as the maximum percentile value of 99,9% (99,9% of all values will be less than the specified value). This is a very important indicator, especially for server drives, since they require predictability, stability and the absence of unexpected freezes in their work. If you know what QoS delays are specified in the drive specification, then you can predict its performance, which is very convenient.
Test Options
Both drives were tested in a test bench that simulates a server. Its characteristics:
- Processor Intel Xeon E5-2620 V4 (8 cores, 2,1 GHz, HT enabled)
- 32 GB storage
- Supermicro X10SRi-F motherboard (1x socket R3, Intel C612)
- CentOS Linux 7.6.1810
- FIO version 3.14 was used to generate the load
And once again about which SSD drives were tested:
- Kingston DC500R 960GB (SEDC500R960G)
- Firmware: SCEKJ2.3
- Size: 960 bytes
- Kingston DC500M 1920 GB (SEDC500M1920G)
- Firmware: SCEKJ2.3
- Объём: 1 920 383 410 176 байт
Testing technique
Based on the popular test suite , however, testers have made adjustments to make the workloads more approximate to the real use of enterprise SSDs in 2019. In the description of each test, we will note what specifically was changed and why.
I/O Operations Test (IOPS)
This benchmark measures IOPS for various block sizes (1024K, 128K, 64K, 32K, 16K, 8K, 4K, 0,5K) and random access with various read/read ratios. record (100/0, 95/5, 65/35, 50/50, 35/65, 5/95, 0/100). Truesystems experts used the following test parameters: 16 threads with a queue depth of 8. At the same time, a 0,5 KB block (512 bytes) was not run at all, since its size is too small to seriously load drives.
Kingston DC500R in IOPS test
Table data:
Kingston DC500M in IOPS test
Table data:
The IOPS test does not involve entering saturation mode, so it is fairly easy to pass. Both drives did a great job, fully complying with the declared factory specifications. The test subjects demonstrated excellent performance in writing in 4 KB blocks: 70 and 88 thousand IOPS. That's great, especially for the reading-oriented Kingston DC500R. As for direct read operations, these SSDs not only exceed their own factory values, but in general are approaching the performance ceiling of the SATA interface.
Throughput test
This test examines throughput for sequential access. That is, both SSDs perform sequential read and write operations in blocks of 1 MB and 128 KB. 8 threads with a queue depth of 16 per thread.
Kingston DC500R:
- 128 KB sequential read: 539,81 MB/s
- 128 KB sequential write: 416,16 MB/s
- 1 MB sequential read: 539,98 MB/s
- 1 MB sequential write: 425,18 MB/s
Kingston DC500M:
- 128 KB sequential read: 539,27 MB/s
- 128 KB sequential write: 518,97 MB/s
- 1 MB sequential read: 539,44 MB/s
- 1 MB sequential write: 518,48 MB/s
And here we also see that the sequential read speed of the SSD has crept up to the bandwidth limit of the SATA 3 interface. In general, Kingston drives have no problems with sequential reading.
Sequential write lags a little, which is especially evident in the Kingston DC500R, which belongs to the read intensive class, that is, it is designed for intensive reading. Therefore, the Kingston DC500R in this part of the test gave out values even lower than the declared ones. But Truesystems experts believe that for a drive that is not designed for such loads at all (recall that the resource of the DC500R is 0,5 DWPD), these 400-odd MB / s can still be considered a good result.
Delay test
As we have already noted, this is the most important test for corporate drives. After all, it can be used to determine what problems arise during long daily use of an SSD drive. The standard SNIA PTS test measures the average and maximum latency for various block sizes (8 KB, 4 KB, 0,5 KB) and read/write ratios (100/0, 65/35, 0/100) at the minimum queue depth (1 thread with QD=1). However, Truesystems decided to seriously improve it to get more realistic values:
- Excluded block 0,5 KB;
- Instead of a single-threaded load with queues 1 and 32, the load varies in number of threads (1, 2, 4) and queue depth (1, 2, 4, 8, 16, 32);
- Instead of 65/35 ratio, 70/30 is used as more realistic;
- Not only average and maximum values are given, but also percentiles 99%, 99,9%;
- for the selected value of the number of threads, graphs of latency (99%, 99,9% and average) versus IOPS are plotted for all blocks and read/write ratios.
The data were averaged over four of the 25 rounds of 35 seconds (5 “warm-up” + 30-second load) each. For the graphs, Truesystems chose a series of values with a queue depth of 1 to 32 with 1-4 threads. This was done in order to evaluate the performance of drives taking into account the delay, that is, the most realistic indicator.
Average latency metrics:
This graph clearly demonstrates the difference between the DC500R and DC500M models. Kingston DC500R is designed for intensive read operations, so the number of write operations practically does not grow with increasing load, remaining at the level of 25.
When looking at a mixed load (70% write and 30% read), the difference between the DC500R and DC500M is also noticeable. If you take a load corresponding to a delay of 400 microseconds, you can see that the versatile DC500M has three times the performance. This is also quite natural and stems from the characteristics of the drives.
An interesting detail is that the DC500M outperforms the DC500R even at 100% reading, delivering lower latency for the same amount of IOPS. The difference is small, but very interesting.
99% latency percentile:
99.9% latency percentile:
According to these charts, Truesystems experts verified the accuracy of the declared QoS latency performance. The specifications indicated a figure of 0,5 ms for reading and 2 ms for writing for a 4 KB block with a queue depth of 1. We are proud to report that these figures were confirmed, also by a large margin. Interestingly, the minimum read latency (280-290 µs for DC500R and 250-260 µs for DC500M) is achieved not at QD=1, but at 2-4.
The write latency at QD=1 was 50 µs (such a low latency is due to the fact that at low load the drive's cache is guaranteed to have time to free up, and we always see a delay when writing to the cache). This figure is 40 times lower than the declared value!
Continuous performance test
Another extremely realistic test that examines the change in performance (IOPS and latency) during long intensive work. Random recording in 4 KB blocks for 600 minutes was chosen as a working scenario. The meaning of this test is that under such a load, the SSD drive enters saturation mode, when the controller is constantly engaged in garbage collection to prepare memory blocks free for writing. That is, this is the most exhausting mode - exactly what enterprise-class SSDs face in real servers.
Based on the results of the test, Truesystems received the following performance indicators:
The main result of this part of the test: both Kingston DC500R and Kingston DC500M exceed their own factory values in real operation. When prepared blocks run out, saturation mode begins, Kingston DC500R keeps at 22 IOPS (instead of 000 IOPS). The Kingston DC20M stays in the 000-500 range, although the drive profile claims 77 IOPS. Also, this test clearly shows the difference between the drives: if a high share of write operations is expected in the workflow of the drive, the Kingston DC78M is more than three times more productive (we also remember that the DC000M performed better in latency in read operations).
The delays during constant write operations are shown in the following graph. Median, 99%, 99,9% and 99,99% percentiles.
We see that the latency of both drives increases in proportion to the decrease in performance, without sharp dips and inexplicable peaks. This is very good, since it is precisely predictability that is expected from corporate drives. Truesystems experts emphasize that testing took place in 8 threads with a queue depth of 16 per thread, so it's not the absolute values that matter, but the dynamics. When they tested the DC400, on this test there were strong delays due to the operation of the controller, but on this chart, the Kingston DC500R and Kingston DC500M do not have such problems.
Load Delay Distribution
As a bonus, Truesystems ran the Kingston DC500R and Kingston DC500M through Simplified Test #13 of the SNIA SSS PTS 2.0.1 specification. We studied the distribution of delay under load in the form of a special CBW pattern:
Block sizes:
Load distribution by storage volume:
Read/write ratio: 60/40%.
After secure erase and preload, testers ran 10 60-second rounds of the main test for thread count 1-4 and queue depth 1-32. Based on the results, a distribution histogram was constructed by values from the rounds corresponding to the average performance (IOPS). For both drives, it was achieved with a single thread with a queue depth of 4.
As a result, the following values were obtained:
DC500R: 17949 IOPS @ 594 µs latency
DC500M: 18880 IOPS at 448 µs.
The latency distribution was analyzed separately for reads and writes.
Conclusion
The Truesystems editors concluded that the performance of the Kingston DC500R and Kingston DC500M tests is clearly interpreted as good. Kingston DC500R copes very well with read operations and can be recommended as a professional equipment for the corresponding tasks. For mixed workloads and larger resource requirements, Truesystems recommends the Kingston DC500M. The publication also notes attractive prices for the entire line of Kingston enterprise drives and admits that the transition to TLC 3D-NAND really helped to reduce the price without losing quality. Truesystems experts also liked Kingston's high level of technical support and a five-year warranty on the DC500 series of drives
PS We remind you that .
For more information about Kingston Technology products, please contact .
Source: habr.com