templexxx / ipc

XSI shared memory in Go

ipc

XSI shared memory in Go.

This lib implements basic APIs of shared memory:

1. shmget
2. shmattach
3. shmdetach
4. shmctl rmid

Users should decide which APIs to use, you could find examples in testing.

Some useful ctls:

1. ipcs
2. ipcrm

Performance

NUMA

I think it's important to know some important details about NUMA before using shared memory.

In NUMA, different CPUs may cross nodes to access the same area of memory, it will cost 10x(10ns -> 100ns) than normal memory random access. The default NUMA policy in kernel is:

local allocation

Which means kernel will allocate memory in same node for a process. If the node's memory is full, kernel will try to swap, it may cause "swap insanity". When a process which doesn't need much memory, it works great.

For shared memory, we may want processes which shared memory each other run on same node, we can use this ctl numactl

It's a useful ctl to control numa binds.

e.g.: Run program “myapp” on cpu node 0, using memory on node 0.

numactl --cpunodebind=0 --membind=0 myapp 

e.g.: Run program "myapp" on physical cpu 0,1,2,3,4.

numactl --physcpubind=+0-4 myapp

Warning:

1. Ensure your applications won't use too much memory.

2. If you want bind node after starting, please using cgroup.