Analyzers
Support for Analyzers (Intel VTune™ Profiler, Intel Advisor, Intel Inspector)
Announcements
This community is designed for sharing of public information. Please do not share Intel or third-party confidential information here.
4660 Discussions

How to identify the cause for the high CPI rate

mahesh_w_
Beginner
189 Views

I was trying to identify the reason for the lateness of my program. And I notice that one function has high CPI value (4.5), and it says the reason may

  • Memory stalls 
  • Instruction starvation 
  • Branch misprediction 
  • Long latency instructions 

How can I explore those things using Vtune. Can anyone help me to identify the specific reason for the high CPI? 

I am using vtune 2015 U1 (trial version). and i am a windows user 

0 Kudos
1 Solution
Peter_W_Intel
Employee
189 Views

CPI value is high, this is performance result. You need to identify why by using General Exploration analysis to know detail.

Please see if there is any highlight item (pink color) in bottom-up report, item(s) will be in: 

Filled Pipeline Slots :  Retiring (General Retirement,  Microcode Sequencer), Bad Speculation

Back-end Bound: Memory Bound(L1 Bound, Store Bound, L3 Bound), Core Bound (Port Utilization)

Front-end Bound: Front-end Latency, Front-End Bandwidth.

 

 

View solution in original post

6 Replies
Peter_W_Intel
Employee
190 Views

CPI value is high, this is performance result. You need to identify why by using General Exploration analysis to know detail.

Please see if there is any highlight item (pink color) in bottom-up report, item(s) will be in: 

Filled Pipeline Slots :  Retiring (General Retirement,  Microcode Sequencer), Bad Speculation

Back-end Bound: Memory Bound(L1 Bound, Store Bound, L3 Bound), Core Bound (Port Utilization)

Front-end Bound: Front-end Latency, Front-End Bandwidth.

 

 

mahesh_w_
Beginner
189 Views

@peter,

Thank You.

In the general  exploration following values were high(marked in pink), so that means memory stalls caused the problem? 

Unfilled Pipeline Slots
    memory latency
        LLC miss (0.247)
        LLC Hit (0.262)
        DTLB Overhead (0.264)

Bernard
Black Belt
189 Views

I think that LLC miss rate and DTLB page walking overhead could have contributed to large CPI rate.

Peter_W_Intel
Employee
189 Views

mahesh w.

Please drill down source view to investigate why LLC miss and DTLB page walk overhead, you might change data structure for optimizing memory layout.

mahesh_w_
Beginner
189 Views

@iliyapolak and @Peter 

Thank you for your kind helps. I got the answer I wanted.I can continue From here(probably). 

Bernard
Black Belt
189 Views

@mahesh

You are welcome.

Reply