How to choose high performance virtual machine AWS Gu Geyun ucloud Tencent cloud test big PKPosted by Tolstrup Vincent on February 13th, 2021 Lately, with the rapid development of cloud computing, big data and artificial intelligence, the computing power of data center is also facing higher and higher challenges. When it comes to CPU processor selection in data center, amd has gradually won the favor of main-stream cloud manufacturers because of its strong performance, low power consumption and inexpensive of the most recent generation of EYPC processor. It's reported that AWS, Gu Geyun, ucloud, azure, IBM cloud, Tencent cloud and other cloud vendors have launched cloud hosts predicated on AMD EYPC CPU, which are the pioneers of cloud vendors. In order to facilitate users to intuitively compare the performance differences of AMD virtual machines of different cloud vendors, in early May 2020, CIO time academy launched a test for several main-stream amd virtual machines in the marketplace. The AMD virtual machine in this test provides objective and fair test data under the same test environment so far as possible, and analyzes the performance (CPU, disk, network, etc . ) and cost performance evaluation of AMD virtual machines from different manufacturers from the technical level. 1、 The model selection suggests that the tested amd virtual machine models are ucloud Kuaijie AMD and Tencent cloud SA2 in China, and AWS M5a and Google cloud n2d in foreign countries. In the most popular configuration of high-performance virtual machine, we selected 8-core 32g and 2000g data disks for four cloud vendors. The next figure shows the precise configuration information: Figure 1: information on virtual machine configuration 2. Test method: this test mainly uses unixbenIn this paper, we test the CPU performance, disk performance and Intranet quality of the virtual machine by using CH performance test tool, Pi Pi Pi test, FIO, netperf and other test techniques. Additionally , we ensure that you compare three common application scenarios of nginx, MySQL and Clickhouse. Each test index in this report takes the average value of three tests whilst the final data, and each and every time the virtual machine completes a test, it needs to perform for a lot more than 8 hours before the next test. Figure 2: test details of virtual machine 3. Actual performance investigation in order to show the performance differences of four amd virtual machines more plainly and intuitively, we extracted several important technical indicators in the test outcomes for your reference, including CPU time consumption, disk IOPs, packet forwarding PPS, TPS / QPS and so on. 1 ) CPU performance CPU may be the foundation of the stable operation of the virtual machine os, and the running time of CPU largely determines the performance of the virtual machine. In the same test environment, we use BC command to do floating-point operation. How many digits following the decimal point is 5000. The shorter the test time, the better the performance. Will there be a gap in CPU performance involving the four virtual machines designed with AMD EYPC processors? Let's explore: Figure 3: the common CPU computing time (from short to long) according to the figure above, we are able to intuitively see that when it comes to 8-core CPU, ucloud takes the shortest 16. 54 seconds to calculate pi pi pi, and AWS takes the longest11 seconds, Tencent cloud and Google cloud are similar, 17. 42 seconds and 18. 02 seconds respectively. Compared with Tencent cloud and Google cloud, ucloud can reduce CPU time by 0. 88 seconds to at least one. 48 seconds, and save running time by 4. 57 seconds in contrast to AWS. Ergo, in contrast to the other three, ucloud's amd virtual machine can improve CPU performance by up to 21. 6%. 2. Disk performance in terms of storage performance of virtual machine, IOPs and IO delay are important indicators to measure disk performance. Theoretically, the larger the IOPs, the better, and the low the delay, the better. In this test, the FIO tool we used selected the libaio engine test, the block size was 4K, and the file size generated by the test was 200g. Here, we mainly observe the number of random read / write operations per second and the common latency if the queue depth is 1 * 1 and 7 * 16 respectively. Figure 4: random read / write with queue depth of just one * 1 Figure 5: average delay with queue depth of just one * 1 Figure six: detailed data table with disk random read / write at queue depth of 1 * 1 figure 7: random read / write with queue depth of 7 * 16 figure 8: average delay with queue depth of 7 * 16 Figure 9: random read / write with disk at queue depth of 7 * 16 Step by step data dining table is visible from the info chart. When the queue depth is 1 * 1 and 7 * 16, ucloud's IOPs and average delay have obvious advantagesPotential, as the other three are similar. If the queue depth is 1 * 1, the number of random reads in ucloud is 2. five times that of AWS, which ranks second, and the amount of random writes is 10 times higher than Tencent cloud, which ranks last. When the queue depth is 7 * 16, ucloud's random reading and writing ability remains outstanding, which will be four times that of Tencent cloud, ranking second. AWS and Google cloud are relatively close, that are not ideal. 3. Network performance the number of network packets processed by virtual machine per second may be the core index of Intranet performance. The network situation in Intranet environment will directly affect the interaction between services. In this intranet performance test, we used netperf network performance test tool, through 8 auxiliary machines and 1 test machine, using UDP protocol packet, packet size is 1 byte, in the same case, statistics the amount of packets received by the test machine per second. Next, let us have a look at the performance of these four virtual machines: Figure 10: the typical quantity of packets received per second (from more to less) in the intranet. Generally, the more packets received per second by the virtual machine, the better the quality of the intranet. Based on the test results in the figure above, ucloud receives the absolute most packets per second, about 3. 7864 million, which can be 2. 5 times of Tencent cloud. Google cloud, AWS and Tencent cloud receive about 1 ) 5-1. 8 million packets per second. The number gap just isn't very big, and there is a big room for improvement. 4. Specific scene performance. Nginx test: nginx is a high performanceHigh performance HTTP and reverse proxy web server may be the first range of website portal users. Nginx has got the traits of strong concurrency, therefore it has a quite high requirement on the network capability of the operator's host. According to the quantity of requests processed by nginx per second, we could begin to see the pressure resistance of the server. We use AB pressure test tool to check, AB tool can create multiple concurrent access threads, simulate multiple people to search for a URL address at the same time, its principle resembles CC attack. For that reason under such a higher concurrency test, the more how many requests processed per second, the better the performance, and the less the single request time and test time, the better the performance, which entails that the less the access request loss. Figure 11: average packets received per second of nginx (from more to less) Figure 12: detailed data table of nginx test. The figure above may be the average data obtained by us based on the number of AB pressure test requests of 2 million and the number of concurrent requests of 500. It really is obvious that ucloud has got the most readily useful performance in this pressure test, far exceeding other cloud platforms. In terms of processing requests per second, ucloud is 2 times higher than Tencent cloud, and Google cloud, AWS and Tencent cloud rank second, third and fourth respectively; with regards to single request time and test time, ucloud only is the reason about 1 / 3 of the other three. ·The performance status of MySQL database tested by MySQL can usually reflect the performance of CPU, memory and disk of the server, and in addition reflect the performance huge difference of virtual machine. We use sySbench benchmark tool stores 20 million pieces of data in a dining table in the database, and starts 500 threads at exactly the same time. The results show that the bigger how many requests per second QPS and transactions per second TPS, the better the performance; the smaller the test time and average delay, the better the performance. Figure 13: MySQL QPS comparison (from more to less) Figure 14: MySQL TPS comparison (from more to less) Figure 15: MySQL test step by step data dining table QPS and TPS, as the key indicators to gauge the performance of the database, are also the direct embodiment of the performance of the virtual machine. Based on the test data above, in the MySQL scenario, ucloud's QPS and TPS requests are in front of others, accompanied by Tencent cloud, AWS and Google cloud. Through calculation, the QPS and TPS of ucloud virtual machine are 38% - 70% greater than those of the other three. ·Clickhouse is a column database management system for on-line analytical processing (OLAP). It enables all available resources on the server to process large queries in parallel in the most natural way. At present, Clickhouse has been placed on big data analysis and processing business by many users, therefore the ability to support Clickhouse is also a significant embodiment of the good performance of virtual machine. We install the stand-alone version of CLI on the virtual machineClickhouse 19, do data grouping query on Clickhouse, test data set: ontime, 16g, 183. 95 million rows. Then take the first 10 data, the shorter the final time, the better the performance. Figure 16: Clickhouse test time (from short to long) according to the above figure, ucloud takes the shortest 2. 067 seconds, which can be not far behind Tencent cloud, but Google cloud takes the longest time, and ucloud's final time could be paid off by about 42% compared with Google cloud. 4、 Comprehensive cost performance evaluation · performance score besides the above-mentioned basic performance tests of AMD virtual machine, in order to further show the comprehensive performance of the cloud platform amd virtual machine (such as system call, read-write, process, graphical test, 2D, 3D, operation, etc . ), we use UnixBench tools to test and get an intuitive data for comparison. Figure 17: average performance score of virtual machine (from high to low) according to the test outcomes in the above figure, we could see that under the configuration of 8-core amd virtual machine, ucloud gets the highest performance score of about 6720. 4, while Tencent cloud ranks second with a performance score of 5378. 4, AWS is relatively low, with a performance score of 3732. 8, and ucloud's performance score is all about 25% - 80% higher than the other three, respectively. The benefit is very obvious. ·On the cost entirely comparisonAfter comparing the performance differences of the four amd virtual machines, the next issue that users are most worried about may be the price. In the figure below, we find the monthly price of the same configuration of cloud server on the premise of AMD. All cloud hosts are recorded according to the real-time price on the official website, in which AWS converts the annual payment price in to the monthly unit price (Note: whilst the price of foreign cloud manufacturers is affected by exchange rate fluctuations, only the real-time accounting price during this test period is shown here). Figure 18: comparison of unit price of virtual machine (price from high to low) from the purchase price point of view, we could see that ucloud quick hero amd model includes a great advantage in cost, with a monthly price of only 1960 yuan; Tencent cloud SA2 model's monthly price is 2985. 5 yuan, ranking second, which will be more favorable than foreign Google cloud and AWS, and ucloud's price is lower than domestic Tencent cloud Additionally, it may save yourself about 34% in contrast to the price. One of them, the machine price of AWS M5a is the highest, and the average monthly price will surely cost 12140 yuan (the pre configured IOPs is placed at 20000), which can be about six times of ucloud's amd virtual machine. ·Usually, if the comprehensive performance is directly proportional to the actual price, it is problematic for users to make a choice. How to get higher performance of virtual machine within the same budget is the key. Therefore , in order to more intuitively reflect the relationship involving the price and comprehensive performance of AMD virtual machine, we utilize the cost performance formula: cost performance = performance score / unit price. The bigger the score data, the larger the fee performanceHigh cost performance reflects the better performance of the virtual machine beneath the same price. Figure 19: cost performance ranking of virtual machine (from high to low) Figure 20: step-by-step data dining table of cost performance of virtual machine. In line with the data above, we could obviously see that ucloud gets the highest cost performance of 3. 43, Tencent cloud ranks second with 1 . 8, Google cloud ranks third with - 1, while AWS gets the cheapest performance of 0. 31. It can be seen that ucloud's cost performance can be far better than the other three. According to calculation, ucloud's cost performance is about - 9 times that of Tencent cloud, which ranks second, and even 10 times more than that of foreign AWS. Frankly, at the same price, choosing ucloud's amd virtual machine can gain the greatest performance experience. 5、 Evaluation summary this test report mainly focuses on the comparative analysis of the performance indicators of 8-core 32g amd virtual machines of conventional cloud manufacturers at home and abroad: Tencent cloud SA2 model, ucloud Kuaijie amd model, AWS M5a model and Google cloud n2d model. In line with the comprehensive average data, the four cloud vendors have their own pros and cons when it comes to cost performance, CPU performance, disk performance and Intranet quality. But overall, ucloud virtual machine in this test, the test outcomes are great, users in the actual selection, can consider the above mentioned test outcomes and step-by-step data, and with the actual purchase demand for selection decision. spo to spoto clubLike it? Share it! |