False Sharing

• Bad data or work distribution can lead to false sharing

  • example:
    • processor P0 does only even iterations
    • processor P1 does only odd iterations
    • both processors read and write same cache lines
    • each write invalidates the cache line of the pear processor

• • each processor runs cache coherency protocol on the cache lines, thus the cache lines will ping-pong from one cache to another

for(i=myid; i<n; i+= nthr) {

a[i] = b[i] + c[i];

}

Number of processors

Performance

1

2

...

etc.,

normal behaviour

False Sharing

Previous slide

Next slide

Back to first slide

View graphic version