Hi, all.
I am reading a SIGMOD'21 paper named "Maximizing Persistent Memory Bandwidth
Figure 15(a) in the paper shows that cross-NUMA write is very slow:
1) The peak throughput is low
2) Bandwidth collapses when there are concurrent threads.
Anyone know the root causes of 1) and 2) ?
For 1), the authors say it results from the read-modify-write behavior.
Thanks