Bug 385

Summary: latency calculation has potential performance issues
Product: DPDK Reporter: Stephen Hemminger (stephen)
Component: otherAssignee: reshma pattan (reshma.pattan)
Status: UNCONFIRMED ---    
Severity: minor CC: ajit.khaparde
Priority: Normal    
Version: 19.11   
Target Milestone: ---   
Hardware: All   
OS: All   

Description Stephen Hemminger 2020-01-14 19:44:13 CET 
If anyone is using rte_latencystats, they are in for a big performance hit.

The code in rte_latencystats uses a receive and transmit callback for each
burst. The latency calculation is using floating point (slow) and worse yet there is several floating point divides per packet. A floating point divide is one of the slowest opcodes on any processor.

The code should be changed to keep track of cycles (fixed point 64 bit) and convert the result to floating point when doing update. Or better yet only
use fixed point maths.
Comment 1 Ajit Khaparde 2020-02-05 20:54:28 CET 
Reshma, Can you look at Stephen's recommendations? Thanks