question about stalls in perf

question about stalls in perf

[Yunqi Zhang]

Hi all,

Recently, I'm using perf to do some profiling work on SandyBridge.

And I found two events stalled-cycles-frontend and stalled-cycles-backend
very interesting, while I'm not sure what are their accurate definitions.
So my question is which hardware counters on SandyBridge are used to
calculate these two events and how (an equation would be perfect).
Furthermore, I was wondering if it is possible for someone to tell
me in which file this calculation processes in the source code of perf.

Thanks a lot!

Regards,
Yunqi

Re: question about stalls in perf

[William Cohen]

On 02/13/2013 07:08 PM, Yunqi Zhang wrote:
> Hi all,
> 
> Recently, I'm using perf to do some profiling work on SandyBridge.
> 
> And I found two events stalled-cycles-frontend and stalled-cycles-backend
> very interesting, while I'm not sure what are their accurate definitions.
> So my question is which hardware counters on SandyBridge are used to
> calculate these two events and how (an equation would be perfect).
> Furthermore, I was wondering if it is possible for someone to tell
> me in which file this calculation processes in the source code of perf.
> 
> Thanks a lot!
> 
> Regards,
> Yunqi

Hi Yunqi,

It is probably best to find out which specific code are being used to set up counters for those events.  This can be found around the following line of code in the kernel for sandybridge:

http://lxr.linux.no/#linux+v3.7.7/arch/x86/kernel/cpu/perf_event_intel.c#L2069


2068                /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
2069                intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2070                        X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2071                /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
2072                intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
2073                        X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);

The first event counts the number of cycles no ops are issued to the queue.  The 

The events are described in the Intel® 64 and IA-32 Architectures Software Developer's Manual
Combined Volumes 3A, 3B, and 3C: System Programming Guide, Parts 1 and 2 available and the Architecture Optimization Reference Manual from:

http://www.intel.com/content/www/us/en/processors/architectures-software-developer-manuals.html

Table 19-6 of volume 3 (Non-Architectural Performance Events In the Processor Core Common to 2nd Generation Intel® CoreTM i7-2xxx, Intel® CoreTM i5-2xxx, Intel® CoreTM i3-2xxx Processor Series and Intel® Xeon® Processors E5 Family) describes the event for 0x0e and 0xb1.

Chapter 2.1.1 of the Architecture optimizaiton manual describes the sandybridge pipeline.  And B.3.2 "Hierarchical Top-Down Performance Characterization Methodology and Locating Performance Bottlenecks" in the optimization manual describes front end and back end stalls.

-Will

Re: question about stalls in perf

[Yunqi Zhang]

William Cohen <wcohen <at> redhat.com> writes:
> 
> Hi Yunqi,
> 
> It is probably best to find out which specific code are being used to set up 
counters for those events.  This
> can be found around the following line of code in the kernel for sandybridge:
> 
> http://lxr.linux.no/#linux+v3.7.7/arch/x86/kernel/cpu/perf_event_intel.c#L2069
> 
> 2068                /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
> 2069                
intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
> 2070                        X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, 
.cmask=1);
> 2071                /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
> 2072                
intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
> 2073                        X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, 
.cmask=1);
> 
> The first event counts the number of cycles no ops are issued to the queue.  
The 
> 
> The events are described in the Intel® 64 and IA-32 Architectures Software 
Developer's Manual
> Combined Volumes 3A, 3B, and 3C: System Programming Guide, Parts 1 and 2 
available and the Architecture
> Optimization Reference Manual from:
> 
> http://www.intel.com/content/www/us/en/processors/architectures-software-
developer-manuals.html
> 
> Table 19-6 of volume 3 (Non-Architectural Performance Events In the Processor 
Core Common to 2nd
> Generation Intel® CoreTM i7-2xxx, Intel® CoreTM i5-2xxx, Intel® CoreTM i3-2xxx 
Processor Series
> and Intel® Xeon® Processors E5 Family) describes the event for 0x0e and 0xb1.
> 
> Chapter 2.1.1 of the Architecture optimizaiton manual describes the 
sandybridge pipeline.  And B.3.2
> "Hierarchical Top-Down Performance Characterization Methodology and Locating 
Performance
> Bottlenecks" in the optimization manual describes front end and back end 
stalls.
> 
> -Will
> 

Thanks a lot Will! Happy Valentine's Day!