icinga2 host state for passive host

Question

The CUDA kernel function is used to calculate inner product for two vectors. The kernel used multiple threads to calculate products concurrently, and then uses one thread to calculate inner product. My question is why the result would be correct without __syncthreads().

__global__ void dot( int *a, int *b, int *c, int *dot ){
    int tid = threadIdx.x;
    int i;

    c[tid] = a[tid] * b[tid];
    
    //__syncthreads();
    //need synchronize?? 
    if(tid==0){
        for(i=0; i<N; i++){
            *dot += c[i];
        }
    }
}

Solution

The default check set in generic-host is called hostalive, which will check if the host is pingable in any ways.

You would need to change the host state to a dummy check. Since Icinga2 always wants to know if the host is still alive.

Example:

template Host "non-pingable" {
  import "generic-host"

  check_command = "dummy"
  vars.dummy_state = 0
  vars.dummy_text = "Host can not be pinged, should be up, hopefully..."
}

This will run a very generic check, but setting your host state to UP.

You can also use any other check, like SSH, or a port test.

Comments

[mr.zog]

Why do you expect us to know what a "NVR" is?

[Homer512]

What's the block size?

[paleonix]

Probably because you were lucky. This not not guaranteed to work without synchronization.

[Sebastian]

You could test by putting a if(tid > 0)__nanosleep(1000000); in the beginning to sleep 1ms for all except thread 0. You can find out the order of threads by atomically increasing a counter and storing the current value in a separate memory location depending on thread number. The increasing counter number is the order the threads executed this source code location.

[Maarten Ureel]

OK thanks, I am also looking into the distributed monitoring so that the one host that can be reached from the outside would ping the other hosts.

[Hyunwoo Kim]

If you get correct result consistently, I think it's because your if + for loop can be so slow that every thread can complete 2 reads + 1 multiply + 1 write operation.

[Homer512]

It could also be that only one warp is active and the GPU is an older version without independent thread scheduling, correct?

[Hyunwoo Kim]

@Homer512 I'm not sure... I've started cuda since nvidia GPU supported independent thread scheduling. I haven't used old GPUs and thought about that.