Linux Kernel vs DPDK: HTTP Performance Showdown

Summary

In this technical presentation, Mark Richards compares the HTTP performance of the Linux kernel and DPDK (Data Plane Development Kit). By optimizing both networking stacks, he demonstrates that Linux kernel optimizations can significantly improve performance, narrowing the gap with DPDK. While DPDK delivers superior throughput, the kernel, with fine-tuned adjustments, achieves impressive results, reducing the performance gap from 4.2x to 1.5x. The presentation discusses the challenges and trade-offs of using DPDK, including the lack of network monitoring tools and high CPU utilization, while exploring potential kernel optimizations and future performance gains.

Takeaways

• 😀 The presenter, Mark Richards, compares HTTP performance between the Linux Kernel and DPDK in an AWS cloud environment.
• 😀 Kernel bypass technologies like DPDK enable high-performance networking by bypassing the kernel’s network stack, allowing direct communication between the NIC and the application.
• 😀 DPDK can achieve significant performance gains, particularly for specialized applications, by aggressively polling and dedicating the network interface to a single application.
• 😀 Although DPDK offers high performance, it requires specialized optimizations that may not be practical or beneficial for all use cases.
• 😀 The Linux Kernel, while designed for general-purpose use, can be optimized for extreme networking performance with specific configurations.
• 😀 Key Linux Kernel optimizations for HTTP performance include enabling receive-side scaling, transmit packet steering, perfect locality, and busy polling.
• 😀 With the right adjustments, the Linux Kernel's performance can approach that of kernel bypass solutions like DPDK, with a 103% improvement in the presenter’s test case.
• 😀 Performance gains were achieved in the Linux Kernel by optimizing the network queue, minimizing interrupts, and adjusting socket behavior to favor efficient data transfers.
• 😀 DPDK’s performance is still ahead of an optimized kernel-based solution, with a final throughput of 1.5 million requests per second, compared to 1 million requests per second on the Linux Kernel.
• 😀 Despite DPDK's performance advantages, it has challenges such as high CPU utilization, difficulty troubleshooting, and lack of access to kernel-based monitoring and security tools.
• 😀 The presenter suggests that kernel-based networking can be further optimized, and points to IOU ring as a potential solution for reducing syscall overhead in future experiments.

Q & A

What is the main focus of the presentation?

The main focus of the presentation is comparing the performance of the Linux kernel networking stack and the DPDK (Data Plane Development Kit) networking stack in handling extreme HTTP workloads.

What is DPDK and how does it work?

DPDK is a kernel bypass technology that allows data to bypass the kernel's networking stack and be processed directly in user space, offering higher performance for applications by reducing latency and increasing throughput.

What was the experiment conducted to compare the two networking stacks?

The experiment used a simple HTTP benchmark where a server returns a 'Hello World' JSON object, running tests both with the Linux kernel's networking stack and the DPDK stack, to compare their performance under extreme conditions.

What were the initial performance results for DPDK?

DPDK achieved an impressive 1.19 million requests per second out of the gate, demonstrating a strong performance right from the beginning of the test.

How was the Linux kernel's performance optimized?

The Linux kernel's performance was optimized by disabling speculative execution mitigations, enabling receive-side scaling, transmit packet steering, configuring busy polling, and tuning the application for better data locality and efficiency.

What was the performance of the unoptimized Linux kernel compared to DPDK?

The unoptimized Linux kernel performed around 358,000 requests per second, which, while fast, was significantly lower than DPDK's performance of 1.19 million requests per second.

How did the performance improve after applying optimizations to the Linux kernel?

After optimizations, the Linux kernel's performance increased to 726,000 requests per second, a 103% improvement, but still far behind DPDK.

What role did the application-level optimizations play in improving performance?

Application-level optimizations, such as configuring thread pinning, adjusting timeout values, and optimizing system calls like switching from 'read' to 'receive', contributed to improved performance, ultimately bringing the kernel's performance up to 1 million requests per second.

What were some key challenges with using DPDK?

Key challenges with DPDK include its niche nature, difficulty in finding relevant resources, and the absence of the kernel's networking stack, which is time-tested for security, monitoring, and troubleshooting purposes.

What conclusion did the presenter draw about the potential of optimizing the Linux kernel?

The presenter concluded that while DPDK still maintains a performance lead over the Linux kernel, the gap represents unrealized potential for further kernel optimizations. The kernel could approach DPDK's performance, and ongoing developments like IOU ring may help close the gap.