是否可以在“保留”的cpu核心上运行代码？

Question

Simplified background:

My app runs a lot of tasks. Most of them are CPU intensive.
One task (which is actually a single thread running in a loop, listening to packets from the network), is very much a "realtime" task. To make it more interesting, that thread is native code called using pinvoke .

The problem:

When a lot of traffic is showing, the tasks are working very hard, and all the cores are maxing out. When that happens, the "realtime" thread (which runs on a 100% cpu core), starts to drop packets, because it doesn't get enough cpu time.

The question:

It is possible to somehow "reserve" one core for the "realtime" thread, and limit all the other threads (tasks) to other cores?

• Of course, there are other processes running, consuming CPU time as well, but let's assume they consume little and constant resources.
• This is a real problem that can be solved by "throw more cpu on it"... Not an option...

EDIT - Answering many useful comments:

• We use WinPcap to capture all packets, which can be many (many).
• The "realtime" thread is not really "realtime" (I think "realtime" is for processes - In .net ThreadPriority uses "Normal", "AboveNormal", etc..)
• The "realtime" thread calls WinPcap serially, packet after packet. We suspect that since it is starved enough, it doesn't keep up, and WinPcap's buffer is overflowed.

Answer 1

You can specify an affinity mask for processes and threads. This allows you to prevent the scheduler using specific processors. So you could set the thread affinity mask for one thread to a single processor, and set the thread affinity mask for all other threads to be all processors other than the reserved processor. Since .net insulates you from directly creating system threads it is actually not that easy to impose the necessary masks on all system threads.

Another way to persuade the scheduler to give your critical thread special treatment would be to boost its priority. That is something that you should only do with care and significant forethought. However, it's a much simpler way to make sure that your thread never has to wait for other threads in your process.

Perhaps it would be better for your app to use fewer threads than processors, but create a dedicated thread for the critical work. That way you could avoid any affinity mask or priority skullduggery and let the system's thread scheduler do the work for you

Answer 2

Yes. That concept is called "processor affinity", and exists at both the process level and the thread level. It is a little awkward to invoke, but nothing scary: How Can I Set Processor Affinity in .NET?

Note that you would have to set the affinity for all the threads that you see. However, you are not the only creator of threads - things like the ThreadPool and GC also have some threads. I do not suggest it is a great idea to mess with those, but the ProcessThread class won't actively stop you from doing this.