Performance test: select() vs poll()

Purpose

Determine the performance difference between two file descriptors monitoring methods: select() vs poll().

Specs

Machine: AWS EC2 t2.micro (1 vCPU ‒ Intel Xeon, 1 GB RAM), Amazon Linux AMI 2018.03

Silgy version: 4.1

There were two Silgy apps:

1. Server: Silgy Hello World as it was (with default FD_MON_POLL), and with FD_MON_SELECT set, started by ec2-user
2. Client: perf with 10 silgy_svc processes, started by root

There were 2 tests, both with freshly started server:

1. FD_MON_POLL version queried by 100 batches @ 1000 requests each, 3 series

2. FD_MON_SELECT version queried by 100 batches @ 1000 requests each, 3 series

Results

• FD_MON_POLL

  • from the client's point of view:

    Average = 0.19 ms

    24161 per second
    19196 per second
    19987 per second

  • server's point of view:

    --------------------------------------------------
    Memory: 4 760 kB (4.65 MB / 0.00 GB)
    --------------------------------------------------
    
                req: 300000
            average: 0.100 ms
    connections HWM: 10
    

• FD_MON_SELECT

  • from the client's point of view:

    Average = 0.19 ms

    17441 per second
    18956 per second
    18166 per second

  • server's point of view:

    --------------------------------------------------
    Memory: 4 804 kB (4.69 MB / 0.00 GB)
    --------------------------------------------------
    
                req: 300000
            average: average: 0.124 ms
    connections HWM: 10
    

Conclusion

As expected, poll() allows slightly better performance, as it does not require building FD_SET, nor resetting timeout before each call, as select() does. The difference however is not a huge one, below 10%.