How to enable multi threading on SRS

Question

How to enable multi threading on SRS

it's defaultly using only 2 threads which is not enough

I started SRS, it's defaultly using only 2 threads

Answer 1

First, SRS is a single thread media server, and high performance so you normally don't need multiple threading. Because multiple threading is not a silver bullet not even a copper one, it's the problem itself rather than a solution for problem.

Please just finger out why your SRS server performance is low, rather than switch to multiple threading model. If you just want to be as cool as multiple threading, please don't use SRS, because SRS dedicates for supporting business, not for cool.

Why NOT Multiple Threading?

I don't like multiple threading, because it really causes lots of problems:

1. There is always mutex or lock, which hurt performance when the number of threads increasing, and crashes.
2. It's really difficult to monitor a multiple threading server. For example, if the CPU is 320%, could you tell me whether the server is ok or not?
3. K8s pod or docker container allows you to run a set of pods on a set of CPUs, which is similar technology to multiple threading, but much better and stable.

If this doesn't convience you, please do some research about REDIS which is also single thread server.

Why Single Threading?

It's really simple: Because the bottleneck of media server is not the CPU, but the IO or network, please see State Threads for Internet Applications. This is why SRS uses coroutine of state-threads as network framework.

Say, if you want to support 1000 players, and each views a RTMP stream in 2Mbps, then the outgoing bandwidth is 2Gbps. Does your AWS/TencentCloud server support 2Gbps bandwidth? Or do you want to pay for 2Gbps fee, which is about $0.02 per second, $72 per hour, or $1728 per day?

If yes, SRS only needs about 30~50% CPU to deliver live stream to these clients, so you still don't need a multiple threading server.

Ultimately, it's really easy to monitor SRS, because if CPU >80%, the load is too high and you should do something to scale out your SRS server.

Note: Please note that both SRS and State Threads supports multiple threading, so it's not the limitation of technology.

How to Scale-out SRS Server?

Actually, SRS allows you to scale out by live stream cluster, please read origin cluster and edge cluster, which allows you to build a CDN system.

If you deploy SRS cluster on a single node, then you get all your CPU busy, which is exactly the same or even better than multiple threading server.

This is actually the load balance issue for media server, please read Load Balancing Streaming Servers for detail.

And, we're developing a new cluster for WebRTC and SRT, please see #3138, and we'll release it at SRS 6.0.

Does SRS support Multiple Threading?

Yep! We really did lots of research about multiple threading, see #2188. To be honest, we found it's neither a neccessary nor good architecutre for media server.

However, we have refined the structure of SRS to support multiple threading, so there are two threads in SRS running, not really a single thread.

We might use multiple threading for transcoding and disk writing to solve the blocking issue of disk.

But, this doesn't mean we will change our structure to multiple threading model, there is still one thread for stream processing. Again, we want to build a media server for business not for cool.

What's the Limitation of SRS?

Although there are lots of benifits of single thread and cluster model of SRS, but there are still some limitations you should know.

1. Need more instances to maintain. If deploy SRS manually or by K8s, there will be more instances, each is a container or K8s pod. K8s is very good at managing these pods, but you need to learn K8s. If start SRS by docker, you should design the ports and roles for multiple containers in a host server.
2. The goal of system is not a very large system like YouTube, Twitch, AWS or TencentCloud, which supports billion of streams and viewers. Don't use SRS if you want to build a great and giant system.
3. SRS is not good at create a distributed system cross countries and continents, or powerful transport algorithms to work in even 60% loss network. Don't use SRS if you want to build a huge PaaS system.

SRS is design for tiny or medium media system, supports less than 10k connections, uses CDN or PaaS platform such as AWS or TencentCloud to extends the ability for huge concurrency, works with other client open source projects such as FFmpeg, OBS, VLC, WebRTC etc.

Conclusion

SRS is a single thread media server, but it doesn't mean single thread model is not able to scale out, instead we have cluster solution for live stream, WebRTC and SRT, which is much better than multiple threading model.

SRS has multiple threading support in SRS 5.0, so it's not the limitation of technology, but only because of the design choice and the goal and use scenario of SRS.