Multi-core server throughput performance improvement Test

Source: Internet
Author: User

It is very important to improve the throughput performance of multi-core servers. Because servers are different from desktops and workstations, the channel for dealing with the outside world is the network. No matter what complicated computing operations the processor performs, the results must be returned to the request object through the NIC. In a sense, the NIC is the "throat" component of the server for external contact and communication. To some extent, it determines the overall performance of the server. Our multi-core server throughput improvement test is centered on the network card throughput.

In fact, many users have questions about improving the throughput performance of multi-core servers-is the server Nic really so important? My laptop is also configured with a gigabit Nic! Yes, the current mainstream servers are configured with Gigabit Ethernet cards. From the perspective of indicators, the two NICs have no performance difference, but in the face of big data streams, you will find that the server's network adapter is far from comparable to the desktop Nic.

Two types of applications in the test script for improving the throughput performance of multi-core servers of the computer world lab are directly related to the performance of server NICs. One is static Web pages, and the other is ASP-based Web applications. Both applications share the same feature, that is, high concurrent requests per unit time, and massive data packets can be sent directly to the server through a dedicated test device, the server throughput can be immediately displayed.

In the conventional Windows Server 2008 system test, we found that by default, even the best performing Xeon 5500 Server platform, static response indicators are difficult to surpass 28000TPS, at this time, the processor performance is far from being used. On the surface, the throughput performance of the server seems to be limited by the VCPU. In fact, the bottleneck lies in the network card, because the network card can only deal with one VCPU by default, and the server network card should obviously not have only this strength.

From the information and drivers of NIC chips of various servers, we found that TOE (TCP Offload Engine) is provided in the configuration of NIC of the Next Generation servers. This function allows you to move some work at the network layer to hardware to save CPU resources. You can set whether it is effective for Rx/Tx in advanced options. After enabling this option, we can see that the VCPU resources have been released to a considerable extent, and the access performance of the entire machine has also increased to TPS, even in the face of such a strong data flow, the overall CPU usage of the Xeon 5500 platform is still less than 70%, and the performance is still quite stable. In addition to our excitement, we also realized that there are still processors, and it seems that there is still potential to be tapped into on the NIC.

That's right. Next we will introduce the focus of this test-RSS (Receive-Side Scaling). There are a lot of RSS-enabled NICs, but there are only a few queues that support RSS, we finally saw a maximum of eight queue features on the intel i82576 Nic, it also supports 1/2/4/8 queues and gives us a deeper understanding of the significance of the number of queues for the throughput performance of the next-generation multi-core servers.

The essence of the RSS queue function is to distribute NIC data to parallel protocol stacks running on different vcpus. Therefore, in actual tests, we found that when setting two queues, the server performance has been greatly improved. The 90% TPS performance far exceeds the expectation, and the overall CPU usage of the server also exceeds. With the introduction of four queues, even 16 vcpus on the dual Xeon 5500 platform can mobilize more than 95% of the computing power. At this time, the performance almost reaches the peak-57600TPS, this is also the case when we introduce eight more queues. Therefore, we initially determined that to improve the server's Web access, database access, network game front-end, and other application models that emphasize TPS indicators, the RSS queue is an important setting point. It works with TOE to give full play to the throughput performance of servers in the multi-core era.

  1. Overview of server testing methods in terms of performance and functionality
  2. Brief Introduction to the server testing tool for Performance Benchmark Testing
  3. Use Avalanche for server testing

Contact Us

The content source of this page is from Internet, which doesn't represent Alibaba Cloud's opinion; products and services mentioned on that page don't have any relationship with Alibaba Cloud. If the content of the page makes you feel confusing, please write us an email, we will handle the problem within 5 days after receiving your email.

If you find any instances of plagiarism from the community, please send an email to: info-contact@alibabacloud.com and provide relevant evidence. A staff member will contact you within 5 working days.

A Free Trial That Lets You Build Big!

Start building with 50+ products and up to 12 months usage for Elastic Compute Service

  • Sales Support

    1 on 1 presale consultation

  • After-Sales Support

    24/7 Technical Support 6 Free Tickets per Quarter Faster Response

  • Alibaba Cloud offers highly flexible support services tailored to meet your exact needs.