For the past three months I have been contributing to FRRouting as part of Google Summer of Code 2020 (GSoC 20). I am writing this post for the purpose of final project submission and also it will serve as a summary of my whole GSoC experience.
FRRouting (FRR) is an IP routing protocol suite for Linux and Unix platforms. FRR’s seamless integration with the native Linux/Unix IP networking stacks makes it applicable to a wide variety of use cases including connecting hosts/VMs/containers to the network, advertising network services, LAN switching and routing, Internet access routers, and Internet peering.
My GSoC project “Dataplane Batching” aimed to enhance the communication with the kernel. Here is a short background and the motivation of the project.
Thanks to the FRR design architecture, there is the abstracted queue of dataplane context objects (structures containing all required data for e.g., a route addition) for messages waiting to be sent. They are then dequeued and sent to the appropriate dataplane. They were processed one at a time, so there was room for improvement.
This roughly shows how it worked before:
FRR uses several different methods of the communication with the kernel and some of them (the Linux Netlink) thankfully support batch sending messages. So we could take advantage of that and as a result improve overall processing time. This is what the project generally was about. Basically, the goal was to adjust the data plane layer code to process multiple context objects at a time and then to add support for sending and receiving batches of messages via Netlink. Referring to the picture above, I had to add an ability to pass many messages between dataplane and the kernel in one go.
In this section I will go through the technical aspect of my work done over the summer, so if you are not interested in this topic go ahead and skip it.
I started from changing FRR’s dataplane abstraction code to make it capable of processing multiple context object. Actually, it was quite straightforward. In brief, instead of passing a single context object to the lower layer, a list of contexts is now handed. With that being done, I could add support for batch sending messages in the kernel-facing code. The primary goal was to implement this for the Linux and possibly other OS, but anyway it turned out that, as of now, BSD’s routing sockets do not support batching.
The next step was to create a new API for encoding Netlink messages, as the old one was not too consistent. I reorganized it in a way that all the functions take a context object and encode it to the given buffer. The old API also lacked any awareness of the size of the buffer, so it used an assumption that the message would always fit in the buffer, which is not the safest assumption. As a result, we now have more modular and safer API with well-defined interface. Also, it makes this API easier to be reused by other code.
Then it was time to write core utility functions for batching. Nothing special, but the thing worth mentioning is that there is one global buffer that is directly used for encoding messages. So when we send some updates to the kernel, we do not neither allocate any big chunks of memory nor make unnecessary copies. I know, premature optimization is the root of all evil, but it does not affect the readability of the code that much and it was not hard to write.
Error handling was quite tricky. When we send a batch of messages some of them may fail. We need then to corelate those failed messages with context objects and alert the upper level code about it. The response message has the same sequence number as the request we sent, so by having sequence number mapping we can easily match received messages with context objects.
Last but not least part of my work was to write some tests. The code I wrote is tightly integrated with the existing one, so I decided not to write any unit tests, as they would require many mockups. So at least I tried to make some integration tests, but also it was not that easy. The first issue was that the batching is not deterministic. The producer of context objects (main thread) is independent of the consumer (dataplane thread), so the size of the batch highly depends on the scheduler. The second problem was that there was not any method to make the kernel return error message. It could be done by dataplane plugins, but this would require to incorporate them in to the testing system and that could take some time to do. In the end, I added an ability to dynamically change the size of the buffer used in the batching, so we are able to shrink the buffer and then just send many requests to the kernel using SHARP, a special daemon that provides miscellaneous functionality used for testing FRR.
Considering that I was able to meet all my planned goals and also participate in work on additional issues, I would say I was able to surpass the expectations of the project.
The real objective of the project was to improve performance of the code, but it was not certain at the beginning how it would really speed up processing time. There are many more factors that influence this, so the only way to check it is with performance tests. This was one of the most exciting part of the coding phase.
Here are the results:
This graph shows time in seconds required to install 1M routes and then to delete them. You can see that not only the overall processing time has been reduced, but also the share of the dplane thread is now much smaller. This is mostly because there are now less user to kernel transitions than there were before.
I also did performance analysis using flame graps. The samples were taken while installing 1M routes. The dplane thread is the rightmost column “zebra_dplane”. You can see that the most of work done now is in the kernel, which is a good sign, because there is not much left to optimize.
It is also worth looking at the flame graph taken without batching enhancement, so the differences can be spotted:
If you have troubles in your browser (flame graphs should be interactive), try the direct SVG version: batching,pre-batching.
The following is the chronological list of pull-requests as required for the final project submission:
- zebra: fix netlink batching
- zebra: dataplane batching
- zebra: netlink cleanup
- zebra: prepare dplane for batching
- Clean up the zebra’s Netlink API
I also worked on a few different issues. They were more or less related to my main project, but they concered the same part of FRR.
- zebra: remove fuzzing stuff
- zebra: refactor netlink fuzzing
- ospfd: make proactive ARP configurable
- lib, zebra: add support for sending ARP requests
- zebra: fix the installation of the evpn neighbor update
- zebra: move ip rule installation to use dplane thread
- zebra: Add vrf name and id to debugs
- zebra: make common function for RTM_NEWNEIGH calls
FRRouting is the first open source project I have contributed to. I came across FRR when seeking a GSoC organization. There were several factors I made my decision on. Firstly, I was looking for fast-paced project, so my patches would not sit, age, and go nowhere. It is also important to me to have smooth and efficient communication. On FRR’s Slack I usually got response within several minutes and it was by no means because I was a potential GSoC student back then. Another key factor was that it could be a great opportunity to learn more about computer networks. The project also covered many fields that I am especially interested in such as low-level programming and interactions with the kernel.
What I liked the most about the project is that my work will actually be used. It is not a very big feature, but it modifies the crucial part of the code where most of the communication with the kernel happens. This is really satisfying and motivating especially when realizing that FRR is widely deployed in so many places.
There were some challenges I faced during the program. The biggest one was related to the fact that I had to integrate a new feature into the existing code. I spent a great deal of my work time wondering how to handle some edge case without breaking anything else. In the result, the code was not always that clean and simple as planned, but I am still happy with the results. Other challenge was to deal with the large code base and understand it to some extent so that I could extend its functionalities. On top of that, I had to learn more details about networking in the Linux not only for the project itself, but also to be more proficient using miscellaneous tools.
Overall, Google Summer of Code was a great experience and also a good opportunity to learn how open source communities function. This was nowhere close to what I had worked on before. I hope it will not be my last contribution to a open source project.
I would like to express my special thanks to Stephen Worley and Quentin Young for being the awesome mentors. It was a pleasant collaboration mainly because of their guidance, patience, excellent explanations and quick responses even during the weekends. I am also very grateful to FRRouting community who reviewed my work and provided feedback, especially Mark Stapp and Donald Sharp. The pictures above are taken from Donald’s slides. I also was able to attend this year’s Netdev conference because of his support. Once again a big thank you. And of course many thanks to Google for giving me the opportunity to work on the project.