Over a million developers have joined DZone.
{{announcement.body}}
{{announcement.title}}

Testing and Optimization for the Internet of Vehicles: IoV Series

DZone 's Guide to

Testing and Optimization for the Internet of Vehicles: IoV Series

Cloud-based deployment can resolve issues associated with the traditional IoV architecture.

· IoT Zone ·
Free Resource

In our previous articles, we looked at the challenges of a traditional IoV architecture and discussed how a cloud-based deployment can resolve these issues. In this article, we will discuss testing and performance optimization for key IoV services in the cloud.

Server Load Balancer Instance Selection and Performance Indicators

A performance-sharing Server Load Balancer instance cannot guarantee instance performance indicators, but a performance-guaranteed one can provide guaranteed performance. The performance-guaranteed Server Load Balancer instance is recommended out of consideration for the high concurrency features of the IoV industry. The performance-guaranteed Server Load Balancer instance shows three key performance indicators:

  1. Max Connection: Max Connection defines the maximum number of connections that a Server Load Balancer instance can bear. When the number of connections on the instance exceeds the maximum number, new connection requests are discarded.
  2. Connection Per Second (CPS): CPS defines the speed of connection creation. When the connection creation speed exceeds the specified CPS, new connection requests are discarded.
  3. Query Per Second (QPS): QPS defines the number of HTTP/HTTPS queries (requests) that can be completed in a second. It is a special performance indicator for an L7 listener. When the request speed exceeds the specified QPS, new connection requests are discarded. 1

Note: these types are available on the console, with the highest specification of 1,000,000 connections. For a large scale IoV enterprise with a Server Load Balancer demand for 10,000,000 connections, what should we do? The service is not available on the console, as those instances available on the console are provided to small- and medium-sized enterprise users. Enterprise users with higher performance demands can contact your Alibaba Cloud account manager to customize your Server Load Balancer instances. Server Load Balancer instances supporting as many as a hundred million connections are available.

As performance guaranteed instances are provided with specific SLAs on the maximum number of connections, the number of connections per second, and number of queries per second, tests on these instances are omitted. You can check the performance indicators on CloudMonitor in real time when using these instances.

Elastic Compute Service Instance Selection and Performance Test

You can select the ECS instance type based on the use case. For details, see the recommendations on the official website.

2

Alibaba Cloud Elastic Compute Service (ECS) provides different instance types for different use cases to satisfy diversified user demands.

Considering the high concurrency, high throughput feature of the IoT industry, the computing type (C5) with quad core, 8 GB memory, and cloud SSD system disk is recommended for web frontend application, and the general type (G5) with quad core, 16 GB memory, and cloud SSD system disk is recommended for backend application.

RDS Test and Optimization

The following is a performance test on Alibaba Cloud ApsaraDB for RDS MySQL 5.6.

Test Environment

  1. All tests are performed in Zone B of China East 2 (Shanghai).
  2. ECS series II instances are used in the test.
  3. The instance configuration is octa core, 16 GB memory.
  4. The network type is a classic network.
  5. The image for stress testing is CentOS 7.0 (64-bit).

Test Tools

SysBench is a cross-platform and multithread modular benchmarking tool used to evaluate the performance of core parameters when high-load databases are running in the system. SysBench helps you quickly know the database system performance without complex database benchmark settings and even with no database installed.

Performance Indicators

Transactions per second (TPS) indicates the number of transactions processed by the database per second. The counted transactions are committed transactions.

Queries Per Second (QPS) indicate the number of SQL statements such as INSERT, SELECT, UPDATE, and DELETE that the database runs per second.

Test Results

General

22

Dedicated

3

4

5

MySQL Instance Parameter Optimization Suggestions

Some parameters of ApsaraDB for MySQL instances can be modified on the console. When some important parameters are set inappropriately, the instance performance may be affected or an error may occur. This document provides some suggestions for important parameter optimization. Suggestions highlighted in red are for the optimization of the IoV scenarios featuring high concurrency, large data volume, and many read-write operations.

back_log (Set a large value for high-concurrency scenarios)

Default value: 3000

Symptom: If the value is too small, the following error may occur:

SQLSTATE[HY000] [2002] Connection timed out;

Suggestion: Increase the value of this parameter.

innodb_autoinc_lock_mode (It can avoid the deadlock and improve the performance)

Default value: 1

Symptom: When auto-incrementing table locks are used for loading data through the INSERT ⋯ SELECT statement or REPLACE ⋯ SELECT statement, applications may encounter deadlocks during the concurrent data import process.

Suggestion: We recommend that you change the value of innodb_autoinc_lock_mode to 2 to enable the use of the lightweight mutex lock (only in row mode) for all types of Insert operations. This avoids auto_inc deadlocks while greatly improving the performance of the INSERT ⋯ SELECT statement.

NOTE: If the parameter value is set to 2, you must set the format of binlog to row.

query_cache_size (you are recommended to set this value to 0 for IoV scenarios with many read-write operations.)

Default value: 3145728

Symptom: The database goes through a number of different statuses, including checking query cache for a query, waiting for query cache lock, and storing results in the query cache.

Suggestion: ApsaraDB for RDS disables the query cache by default. If the query cache is enabled for your instance, you can disable it when the preceding problem occurs.

net_write_timeout (This parameter can avoid car data write failure caused by timeout.)

Default value: 60

Symptom: If the parameter value is too small, the client may report an error "the last packet successfully received from the server was milliseconds ago" or "the last packet sent successfully to the server was milliseconds ago."

Suggestion: The default value of this parameter for ApsaraDB for RDS instances is 60s. A small value of net_write_timeout may result in frequent disconnections when the network condition is poor or it takes a long time for the client to process each block. In this case, we recommend that you increase the parameter value.

tmp_table_size (you are recommended to enable this parameter in large memory scenarios to improve the query performance.)

Default value: 2097152

Symptom: SQL execution takes a longer time if temporary tables are used when a complex SQL statement contains GROUP BY and DISTINCT clauses which cannot be optimized through an index.

Suggestion: If the application contains many GROUP BY and DISTINCT clauses and the database has enough memory, you can increase the values of tmp_table_size and max_heap_table_size to improve query performance.

loose_rds_max_tmp_disk_space

Default value: 10737418240

Symptom: If the temporary file size exceeds the value of loose_rds_max_tmp_disk_space, the following error may occur.

The table '/home/mysql/dataxxx/tmp/#sql_2db3_1' is full

Suggestion: Check whether the SQL statements resulting in additional temporary files can be optimized by indexing or other means. If your instance has enough space, you can increase the value of this parameter to guarantee normal execution of SQL statements.

loose_tokudb_buffer_pool_ratio

Default value: 0

Suggestion: If you use the TokuDB storage engine in your ApsaraDB for RDS instance, we recommend that you increase the value of the parameter to improve the access performance of the TokuDB engine table.

loose_max_statement_time

Default value: 0

Symptom: If the query time exceeds the value of this parameter, the following error may occur.

ERROR 3006 (HY000): Query execution was interrupted, max_statement_time exceeded

Suggestion: To control the SQL execution time in the database, enable this parameter. The unit is milliseconds.

loose_rds_threads_running_high_watermark (This parameter can be used for protection purposes in high concurrency scenarios.)

Default value: 50000

Suggestion: This parameter is often used for handling seckilling or highly concurrent requests, offering effective database protection.

Elasticsearch Performance Test and Optimization

Elasticsearch is a Lucene-based search and data analysis tool that provides a distributed service. Elasticsearch is an open-source service conforming to the Apache open-source terms and a mainstream enterprise-level search engine.

Alibaba Cloud Elasticsearch provides the open-source Elasticsearch v5.5.3 and the X-Pack Business Edition for the scenarios such as data analysis and data search. A range of functions such as enterprise-level rights management, security monitoring and alarms, and automatic report generation are built upon open-source Elasticsearch. The following is a test of the read/write performance of Alibaba Cloud Elasticsearch.

Write Performance

Test Environment

Clusters of three Elasticsearch 5.5.3 instances, respectively of types dual core 4 GB, quad core 16 GB, and 16 core 64 GB, 1 TB SSD instances, are used in stressing test. esrally is used for stressing testing.

Rest API

There are two types of stressing test data:

  1. esrally official data geonames 3.3 GB, with a single 311 B file.
  2. Service simulating log data log 80 GB, with a single 1432 B file.

Test Results

  1. The number of shards is 6 for all clusters in the test.
  2. The CPU usage, load, and IOUTIL are all approximate values.
  3. The log indicator is tuned based on the asynchronous translog/merge policy for the batch log write test scenarios.

6

Write Performance Indicator Comparison

Impact of the machine on the write performance

  1. The number of shards is 6 for all clusters in the test.
  2. The number of copies is 0.
  3. The SSD capacity is 1 TB in the cloud for all instances.

7

Impact of the Number of Copies on the Performance Indicators

  1. The number of shards is 6 for all clusters in the test.
  2. The data samples are all logs in the test.
  3. The SSD capacity is 1 TB in the cloud for all instances.

8

Query Performance

Test Environment

  1. Clusters of three Elasticsearch 5.5.3 instances, respectively of types dual core 4 GB, quad core 16 GB, and 16 core 64 GB, 1 TB SSD instances, are used in stressing test.
  2. esrally stressing test is used.
  3. esrally official data geonames 3.3 GB, with a single file sized 311 B, is used.

Test Result

  1. The number of shards is 6 for all clusters in the test.
  2. The number of copies is 0.
  3. The query types are term and phrase.

9

Machine Qps Cpu Load JVM Memory 90th Percentile Service Time

2 core 12378 89% 3.98 47% 17.6141ms

4 core 23498 93% 4.63 48% 20.3789ms

ApsaraDB for HBase Performance Test and Optimization

ApsaraDB for HBase is a stable, reliable, and elastically scalable distributed database service based on NoSQL, compatible with the open-source HBase protocol. The functions available and to be available consist of security, public network access, HBase on OSS, backup, and recovery, and cold/hot data separation. The supported storage media are ultra cloud disk, cloud SSD, and local disks. HBase may run on a single host or distributed model.

10

11

Parameter Optimization Recommendation for an HBase Client Instance

  1. Encapsulate the HBase client in Htable. During a request for the HBase, the AutoFlush parameter (true by default) can be set to false for Htable, indicating that flush is carried out only when the local cache of Htable is full. It improves the write speed perceived by the client.
  2. The number of scan caching parameters can also be configured during encapsulation of the HBase client, to define how many data entries are cached each time from the server for scanning by the user. The default value is 1. You are recommended not to set the value too great if the single entry data value is great.
  3. Scan attribute selection is recommended for scan operations. If only one column family is to be scanned, only the data under the column family is operated during the return. If multiple column families are selected for scanning, many unnecessary column families are returned, generating unnecessary communication traffic.
  4. The WAL flag can be set to false in a put operation to enhance the image performance when the security requirement is not high and inconsistency is allowed for unimportant data. In this case, WAL is not written first during a data write operation, improving its performance.
  5. Enable the bloomfilter. If the bloomfilter is enabled for a read operation, the read performance is improved as a result of exchange of space for write operation time.

HiTSDB Performance Test and Optimization

HiTSDB Edge is a high performance time series database running close to the client.

Test Model

The data model used in the test is based on metrics generated randomly. Timelines metric and tagKey, and tagValue consist of tags, that is, a character string in a fixed length (10 bytes) and an index. Timelines metric and tagKey, and tagValue with real meaning are not used as they cannot cover various real scenarios and the test purpose is to compare HiTSDB with other time series databases horizontally. Testing with timelines without real meanings can generate reliable test results when the test data and software and hardware conditions are maintained consistent for these databases. 8 byte long integer time stamps and 8 byte dual precision integer values are used. The data sample is as follows:

12

Description of the Statistical Result

Each time before testing, the database service is restarted to avoid impact of the previous operation results on the cache.

Transactions per second (TPS) is the number of data point operations completed in a time unit.

Each time after a write operation, the timeline is manually selected at random for data query, and the data write consistency is verified based on the granularity. For a HiTSDB write operation, the SDK callback is registered, so that the operation failure or success is perceived.

In addition, to avoid the impact of the test data, data is cleared before each test.

The number of timelines is verified on the HiTSDB internal interface/API/TScount during each test. The increase of the timeline number meets with the expectation. Single-node comparison tests for HiTSDB and other databases

Write Performance

One write request for a database may include one or multiple data points. Generally, the more data points a request contains, the higher is the write performance. In the following test, P/R, points/request, indicates the number of data points in a request.

Write performance of a single client under 500 points per request

13

Write performance of a single client under 1 million timelines

Image title

Client CPU Usage

15

Based on the analysis of the result in the table below, the write TPS decreases when the number of timelines increases on the HiTSDB. The write TPS is about 50,000 when the number of timelines reaches 1 million in 6xlarge, which is close to the 60,000 TPS performance released on Alibaba Cloud official website.

Write performance of dual clients under 500 points per request

16

Write performance of dual clients under 1 million timelines

Image title

Client CPU Usage

18

Write performance of four clients under 500 points per request

19

Write performance of four clients under 1 million timelines

20

Client CPU Usage

21

According to the results of these three test groups, the write TPS of HiTSDB decreases when the number of timelines increases. In the 10,000 and 100,000 timeline number levels, the TPS does not change greatly with the number of clients. In the 1000,000 timeline number level, the TPS increases when the number of clients increases from 1 to 2, and does not increase when the number of clients increases from 2 to 4. Therefore, it can be considered that the HiTSDB write TPS peak is about 150,000 in 6xlarge with 1000,000 timelines. The data is cleared before each test. The write TPS decreases (to about 50,000 to 60,000 according to a test) when the timelines increase in the environment. Therefore, 150,000 is a peak value.

Optimization Recommendation

  1. Write optimization Multiple write data entries for one batch
  2. Query optimization To query data in a long time range, split the query by hour
Topics:
architecture ,database ,trends ,migration ,iov ,iot ,test environment ,mysql ,performance

Published at DZone with permission of

Opinions expressed by DZone contributors are their own.

{{ parent.title || parent.header.title}}

{{ parent.tldr }}

{{ parent.urlSource.name }}