Performance Fluctuates When the Number of Queries Exceeds 1000

Symptom

When the number of queries is greater than 100, the performance fluctuates.

Possible Cause

During concurrent CPU processing on the host, the CPU is scheduled to a non-affinity CPU core. As a result, the time consumption increases.

Solution

You need to bind cores to the retrieval application as follows.

1. Obtain the NUMA node information. As shown in Figure 1, you can see that the queried NPU belongs to NUMA node 0.
   Figure 1 NUMA node information
   
2. Run the lscpu command to view the CPU core information on NUMA node 0. As shown in Figure 2, the CPU core of NUMA node 0 is 0-13, 28-41.
   Figure 2 CPU core information
   
3. Bind the current retrieval application to the confirmed CPU.

   taskset -c 0-13,28-41 ./mxIndexApp

   mxIndexApp indicates the retrieval application to be bound. Replace it with the actual application name.