如何将perf.data转换为文本/CSV？提取1ms采样的硬件计数器数据

Got it, let's tackle how to pull that 1ms interval hardware counter data out of your perf.data file. Since perf stat doesn't support intervals smaller than 10ms, working directly with the perf.data from your 1ms-sampled perf record run is the right approach. Here are the most straightforward methods:

1. Use perf script (Perf's Built-in Tool) #

perf script is designed to parse and output raw data from perf.data in human-readable or machine-friendly formats. It’s perfect for extracting timestamped counter values.

Basic Raw Data Export #

To get each sample’s timestamp, event name, and counter value (period), run:

perf script -F time,event,period -i perf.data > raw_counter_data.csv

• -F time,event,period: Specifies which fields to output (timestamp in seconds, the hardware event you sampled, and the counter’s raw value for that sample).
• -i perf.data: Explicitly tells perf script to use your existing data file (optional if it’s the default perf.data).

If you need additional context (like which CPU/process the sample came from), add more fields to the -F flag. Run perf script -F list to see all available fields. For example:

perf script -F time,cpu,pid,comm,event,period -i perf.data > raw_data_with_context.csv

Aggregate to 1ms Intervals #

If you want to group multiple samples that fall into the same 1ms window (to account for minor timing jitter), use a simple awk script to sum values per millisecond.

Create a file named aggregate_1ms.awk with this content:

BEGIN {
    prev_ms = 0
    total = 0
    print "timestamp_ms,counter_total"
}
{
    curr_ms = int($1 * 1000)  # Convert timestamp from seconds to milliseconds
    if (curr_ms != prev_ms) {
        if (prev_ms != 0) {
            print prev_ms "," total
        }
        prev_ms = curr_ms
        total = $3
    } else {
        total += $3
    }
}
END {
    print prev_ms "," total
}

Then run:

perf script -F time,period -i perf.data | awk -f aggregate_1ms.awk > aggregated_1ms_counter.csv

This outputs a CSV where each line represents the total counter value for a single millisecond interval.

2. Custom Scripting for Advanced Needs #

If you need more control (like filtering specific events or complex data transformations), use perf script with a custom Python/Perl script. Here’s a minimal Python snippet to parse and structure the data:

import sys
import csv

writer = csv.writer(sys.stdout)
writer.writerow(["timestamp_ms", "event", "period"])

for line in sys.stdin:
    parts = line.strip().split()
    if len(parts) < 3:
        continue
    timestamp_sec = float(parts[0])
    timestamp_ms = int(timestamp_sec * 1000)
    event = parts[1]
    period = int(parts[2])
    writer.writerow([timestamp_ms, event, period])

Run it like this:

perf script -F time,event,period -i perf.data | python parse_perf.py > parsed_counter_data.csv

Quick Check on Your Initial perf record Command #

Just to confirm, you should have run something like this to generate your perf.data with 1ms sampling:

perf record -e cycles -i 1ms -o perf.data -- ./your_target_program

Replace cycles with whatever hardware event you’re tracking (e.g., instructions, cache-misses).

内容的提问来源于stack exchange，提问作者onkar.omr