that caused the PMU overflow will have occurred some number of cycles, typically in the low tens, before the address being sampled. Thus, the address recorded might or might not point to the instruction causing the event, depending on pipeline stalls.

The latency between the event triggering the sample and the actual sample is not a problem if you are using fprof to find hot spots in your application. It is only an issue if you try to use fprof to find particular instructions that cause the events recorded by the PMU, in which case you must take the latency into account.

icache Measurement Report Description

With the icache measurement, produced by the icache measurement configuration file, HP Caliper measures and reports on instruction cache metrics. This measurement is similar to the dcache measurement.

The report shows two levels of information:

•Exact counts of instruction cache metrics summed across the entire run of an application

•Sampled instruction cache metrics that are associated with particular locations in the application

The report shows measured data by thread, load module, function, statement, cache line, and instruction.

Command-line options allow you to control the amount of data reported, how the data are sorted, and the number of statements and instructions reported for each sampled program location.

Example Command Line for Text Report

$ caliper icache -o reports/icachem.txt ./matmul

Example Command Line for CSV Report

$ caliper icache --csv csvout ./matmul

icache Metrics Summed for Entire Run

This section describes the metrics summed over the entire run of your application under HP Caliper.

Metrics for Integrity Servers Itanium 2 Systems

L1I_READS

Provides information about the number of demand fetch

 

reads, that is, all accesses regardless of hit or miss, to the

 

L1 instruction cache (32-byte chunks).

 

If demand fetches have an L1 instruction TLB miss, have an

 

L1 instruction cache miss, and collide with a fill-recirculate

 

to the instruction cache, they are not counted in this

 

measurement even though they are counted in

 

L2_INST_DEMAND_READS.

L2_INST_DEMAND_READS

Number of instruction requests to L2 due to L1 instruction

 

demand fetch misses. This event counts the number of

 

demand fetches that miss both the L1 instruction cache and

 

the ISB regardless of whether they hit or miss in the RAB.

 

If a demand fetch does not have an L1 instruction TLB miss,

 

L2_INST_DEMAND_READS and L1_READS line up in time.

 

If a demand fetch does not have an L2 instruction TLB miss,

 

L2_INST_DEMAND_READS follows L1I_READS by 3-4 clocks

 

(unless a flushed iwalk is pending ahead of it, which will

 

increase the delay until the pending iwalk is finished).

 

If demand fetch has an L2 instruction TLB miss, the skew

 

between L2_INST_DEMAND_READS and L1I_READS is not

 

deterministic.

204 Descriptions of Measurement Reports