This topic describes how to use tools to test the performance of the ECS instance CPU, memory, network, and volume. This allows you to learn the ECS instance performance based on the data from test results.

Test principles

For the smooth operation of your business, you need to comprehensively test the performance of your ECS instance that hosts the services. Make sure that you follow the principles below during performance tests:

Objective: Clearly specify the specifications of the ECS instances to be tested (of CPU cores, CPU frequency, memory, storage, network bandwidth, operating system, and other components) and testing methods (such as deployment plans, instance numbers, and service load). This helps generate a fair and objective test result.

Reasonable: Run the tests based on your needs and actual scenarios.

Accurate: Select mature, stable, and authoritative testing tools and test cases. This allows the test results to reflect the ECS instance performance accurately.

Testing tools and key metrics

Type	Test tool	Test metrics
CPU	UnixBench	Single-core benchmark score
	UnixBench			Multi-core benchmark score
		SuperPI	Calculation duration
		Sysbench	Average number of events completed per second
Memory	Stream	Bandwidth
		MLC	Latency
		Sysbench	Bandwidth
Network	iPerfs	bps, pps
			Pktgen
			Netperf
Volume	FIO	I/O latency
	FIO			IOPS
		dd	Bandwidth
		Sysbench	IOPS, bandwidth

Test environment

Prepare the environment

Create a VPC. Recommended region: China (Beijing).

In the created VPC, create two ECS instances with the same specification. Then, associate an EIP with one of the instances and mount a data volume to it (recommended capacity: at least 2 TB).

Software version

ECS Instance operating system: In the example provided in this topic, CentOS 8.3 is used.

Test CPU performance

CPU single-core/multi-core performance testing via UnixBench

Calculation duration testing via SuperPI

CPU load testing via Sysbench

Test description

UnixBench is an open-source tool for testing the basic performance of UNIX systems. It is also commonly used for testing ECS instance performance. You can run various tests via UnixBench to assess system performance from different aspects. Each test result is compared against a baseline value to generate an index score. Then, the index scores of all test items are combined to generate an overall test score.

Procedure

Log in to the ECS instance to be tested.

Execute the following commands to download and install UnixBench.

wget https://github.com/kdlucas/byte-unixbench/archive/v5.1.3.tar.gz
tar -zxvf v5.1.3.tar.gz
cd byte-unixbench-5.1.3/UnixBench
make

Execute the following command to run the single-core test.

./Run -c 1

The execution process may take longer, so please wait patiently. The test score will be displayed as shown in the following picture. Refer to the values of "INDEX" for the result. The higher the values, the better the performance.

Execute the following commands to run the multi-core test.

# Obtain the number of CPU threads (core)
cat /proc/cpuinfo| grep "processor"| wc -l
./Run -c {core}

The execution process may take longer, so please wait patiently. The test score will be displayed as shown in the following picture. Refer to the values of "INDEX" for the result. The higher the values, the better the performance.

Test memory performance

Test bandwidth via Stream

Test latency via MLC

Test Memory load via Sysbench

Test description

Stream is an industry-recognized tool for testing memory performance comprehensively. Supported operations include Copy, Scale, Add, and Triad.

For more information, visit the Stream official website and refer to Stream user guide.

Procedure

Execute the following commands to download and install Stream.

wget https://www.cs.virginia.edu/stream/FTP/Code/stream.c
yum -y install gcc

Edit the test script.

gcc -O -fopenmp -mcmodel=medium -DSTREAM_ARRAY_SIZE=200000000 -DNTIMES=100 stream.c -o stream_omp_exe

Parameter	Description
STREAM_ARRAY_SIZE	Specifies the array size used for testing. Default value: 10,000,000. Stream creates three arrays of the size value specified in the STREAM_ARRAY_SIZE parameter. Default element type: double. For reference, the memory used in the script above is 3 × STREAM_ARRAY_SIZE × sizeof(double) = 4.5 GiB. You can modify the value based on your needs.
NTIMES	Specifies the number of runs for the test case. The final result will be the best one among all runs. Default value: 10. Typically, you do not need to change this value.

Execute the following commands to run the bandwidth test.

# On the same NUMA node
numactl -N 0 -m 0 ./stream_omp_exe
# Cross-NUMA nodes
numactl -N 1 -m 0 ./stream_omp_exe

Parameter	Description
numactl	Specifies the NUMA policy for process and memory allocation. For example, -N 0 -m 0 means using the CPU and the memory of NUMA node 0; -N 1 -m 0 means using the CPU of NUMA node 1 and the memory of NUMA node 0.

After execution, the final result score will be displayed as shown below. Refer to the values of "Best Rate MB/s". The higher the values, the better the performance.

Test network performance

Using iPerf3 to test network performance

Using Pktgen to test network performance

Using Netperf to test network performance

Test volume performance

Testing storage performance

Using Sysbench to test FileIO load

Parameter	Description
--latency_matrix	Prints the matrix of intra-NUMA and cross-NUMA memory latencies
-e	Outputs test results based on original prefetcher settings
-r	Outputs the latency result obtained from random accesses

Elastic Compute Service