Skip to content

Commit

Permalink
Merge pull request #61 from converged-computing/cleanup-osu-parsing
Browse files Browse the repository at this point in the history
add support for more specific osu parsing
  • Loading branch information
vsoch authored Sep 17, 2023
2 parents 6368608 + bbbd5a5 commit 24980db
Show file tree
Hide file tree
Showing 7 changed files with 295 additions and 80 deletions.
2 changes: 2 additions & 0 deletions examples/python/network-osu-benchmark/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -40,3 +40,5 @@ $ python run-metric-multiple.py --iter 5
![img/multiple/OSU-MPI_Put-Latency-Test-v5.8.png](img/multiple/OSU-MPI_Put-Latency-Test-v5.8.png)


Note that if you add the option `timed: true` this will add extra 'wrapper' times around the above,
and an extra data output for it in your results.
2 changes: 2 additions & 0 deletions sdk/python/v1alpha1/CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,8 @@ and **Merged pull requests**. Critical items to know are:
The versions coincide with releases on pip. Only major versions will be released as tags on Github.

## [0.0.x](https://github.com/converged-computing/metrics-operator/tree/main) (0.0.x)
- More specific parsing / control for OSU benchmarks (0.0.21)
- Support for OSU benchmark parsing with timed wrappers (0.0.2)
- Allow getting raw logs for any metric (without parser) (0.0.19)
- Refactor of structure of Operator and addition of metrics (0.0.18)
- Add wait for delete function to python parser (0.0.17)
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,2 @@
from .netmark import network_netmark
from .osu_benchmark import network_osu_benchmark
Original file line number Diff line number Diff line change
@@ -1,85 +1,7 @@
# Copyright 2023 Lawrence Livermore National Security, LLC
# (c.f. AUTHORS, NOTICE.LLNS, COPYING)


from .base import MetricBase


class network_osu_benchmark(MetricBase):
"""
Parse the OSU benchmarks output into data!
"""

container_name = "launcher"

@property
def pod_prefix(self):
return f"{self.name}-l-0"

def parse_row(self, row):
"""
Given a row of two values with spaces, parse.
"""
row = row.split(" ", 1)
return [x.strip() for x in row]

def parse_log(self, lines):
"""
Given lines of output, parse and return json
"""
# Get the log metadata
metadata = self.get_log_metadata(lines)

# Split lines by separator
results = []
sections = self.get_log_sections(lines)
for section in sections:
if not section.strip():
continue
section = section.split("\n")
section = [x.strip() for x in section if x.strip()]

try:
datum = self.parse_benchmark_section(section)
except Exception:
print(f"Issue parsing section {section}")
continue
results.append(datum)

return {"data": results, "metadata": metadata, "spec": self.spec}

def parse_benchmark_section(self, section):
"""
A wrapper for parsing in case there is an error we can catch!
"""
# Command is the first entry
command = section.pop(0)

# Each section has some number of header lines (with #)
header = []
while section and section[0].startswith("#"):
header.append(section.pop(0))

# Last row of the header are the column names
columns = header.pop()
columns = columns.replace("# ", "").strip()
columns = self.parse_row(columns)

# The remainder is data, again always two points
data = []
for line in section:
if not line:
continue
row = self.parse_row(line)
row = [float(x) for x in row]
data.append(row)

return {
"matrix": data,
"columns": columns,
"header": header,
"command": command,
}
from metricsoperator.metrics.base import MetricBase


class network_netmark(MetricBase):
Expand Down
286 changes: 286 additions & 0 deletions sdk/python/v1alpha1/metricsoperator/metrics/network/osu_benchmark.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,286 @@
# Copyright 2023 Lawrence Livermore National Security, LLC
# (c.f. AUTHORS, NOTICE.LLNS, COPYING)

import os
import re

from metricsoperator.metrics.base import MetricBase

# prepare more consistent / formatted columns
latency_size_header = ["Size", "Latency(us)"]
average_latency_header = ["Size", "Avg Latency(us)"]
bandwidth_size_header = ["Size", "Bandwidth (MB/s)"]
column_lookup = {
"osu_ibarrier": ["Overall(us)", "Compute(us)", "Pure Comm.(us)", "Overlap(%)"],
"osu_barrier": ["Avg Latency(us)"],
"osu_mbw_mr": ["Size", "MB/s", "Messages/s"],
"osu_init": ["nprocs", "min", "max", "avg"],
"osu_multi_lat": latency_size_header,
"osu_get_acc_latency": latency_size_header,
"osu_latency": latency_size_header,
"osu_fop_latency": latency_size_header,
"osu_put_latency": latency_size_header,
"osu_get_latency": latency_size_header,
"osu_acc_latency": latency_size_header,
"osu_mt_latency": latency_size_header,
"osu_get_latency": latency_size_header,
"osu_cas_latency": latency_size_header,
"osu_latency_mp": latency_size_header,
"osu_latency_mt": latency_size_header,
"osu_bibw": bandwidth_size_header,
"osu_get_bw": bandwidth_size_header,
"osu_put_bw": bandwidth_size_header,
"osu_put_bibw": bandwidth_size_header,
"osu_bw": bandwidth_size_header,
"osu_allgather": average_latency_header,
"osu_allreduce": average_latency_header,
}


def parse_commented_header(section):
"""
Parse a commented (#) header
"""
# Each section has some number of header lines (with #)
header = []
while section and section[0].startswith("#"):
header.append(section.pop(0))
return header


def parse_row(row):
"""
Given a row of two values with spaces, parse.
"""
row = row.split(" ", 2)
return [x.strip() for x in row if x]


def parse_multi_value_row(row):
"""
Parse a row with multiple values.
"""
row = row.split(" ")
return [x.strip() for x in row if x]


def parse_hello_section(section):
"""
Parse the osu hello section
This only has a print line for output
"""
# Command is the first entry
command = section.pop(0)
header = parse_commented_header(section)

# The next row is just a print line
message = section.pop(0)
timed = parse_timed_section(section)
result = {
"matrix": [[message]],
"columns": ["message"],
"header": header,
"command": command,
}
if timed:
result["timed"] = timed
return result


def parse_init_section(section):
"""
Parse the osu init section
This section has one column, and all the values there!
"""
# Command is the first entry
command = section.pop(0)
header = parse_commented_header(section)

# The next row has all the data!
row = section.pop(0)
values = [x.strip() for x in row.split(",")]
data = {}
for entry in values:
field, value = [x.strip() for x in entry.split(":")]
# Do we have a unit?
unit = ""
if " " in value:
value, unit = [x.strip() for x in value.split(" ")]
if unit:
field = f"{field}-{unit}"
data[field] = float(value)

# If we have additional sections (possibly with times)
timed = parse_timed_section(section)
result = {
"matrix": [list(data.values())],
"columns": list(data.keys()),
"header": header,
"command": command,
}
if timed:
result["timed"] = timed
return result


def parse_timed_section(section):
"""
If the remainder is wrapped in time, parse it.
"""
timed = {}
for line in section:
if line and re.search("^(real|user|sys\t)", line):
time_type, ts = line.strip().split("\t")
timed[time_type] = ts
return timed


def parse_barrier_section(section):
"""
Parse a barrier section.
This section is unique in that it has two columns
but the header is not preceded with a #
"""
# Command is the first entry
command = section.pop(0)
header = parse_commented_header(section)

# The columns are the last row of the header
section.pop(0)
result = parse_value_matrix(section)
result.update({"header": header, "command": command})
return result


def parse_multi_section(section):
"""
Parse a multi-value section.
This section has standard format, but >2 values in the matrix
"""
# Command is the first entry
command = section.pop(0)
header = parse_commented_header(section)

# The columns are the last row of the header
header.pop()
result = parse_value_matrix(section)
result.update({"header": header, "command": command})
return result


def parse_value_matrix(section):
"""
Parse a matrix of times
"""
# The remainder is data, again always two points
# If there are real / user / sys at the end, we ran with timed:true
data = []
timed = {}
for line in section:
if not line:
continue

# We found a time
if re.search("^(real|user|sys\t)", line):
time_type, ts = line.strip().split("\t")
timed[time_type] = ts
continue

row = parse_multi_value_row(line)
row = [float(x) for x in row]
data.append(row)

result = {"matrix": data}
if timed:
result["timed"] = timed
return result


def run_parsing_function(section):
"""
Parsing functions for different sections
"""
# The command is the first line
command = os.path.basename(section[0])
result = None
if command in ["osu_ibarrier"]:
result = parse_barrier_section(section)
elif command in [
"osu_bw",
"osu_bibw",
"osu_barrier",
"osu_get_bw",
"osu_put_bw",
"osu_put_bibw",
"osu_mbw_mr",
"osu_multi_lat",
"osu_allgather",
"osu_latency",
"osu_cas_latency",
"osu_put_latency",
"osu_get_latency",
"osu_latency_mp",
"osu_latency_mt",
"osu_fop_latency",
"osu_acc_latency",
"osu_get_acc_latency",
"osu_allreduce",
]:
result = parse_multi_section(section)

# Special snowflakes
elif command == "osu_init":
result = parse_init_section(section)

# This only potentially has a time :)
elif command == "osu_hello":
result = parse_hello_section(section)

# Columns aren't predictible, so we ensure they are more consistent this way
# Some parsers do their own columns
if "columns" not in result:
result["columns"] = column_lookup[command]

return result


class network_osu_benchmark(MetricBase):
"""
Parse the OSU benchmarks output into data!
For pair to pair we had a common format (two values with size and another field)
but adding on the multi- benchmarks, we have a new challenge that there is slight
variance in format, so I needed to extend the class to be more specific to parsing.
"""

container_name = "launcher"

@property
def pod_prefix(self):
return f"{self.name}-l-0"

def parse_log(self, lines):
"""
Given lines of output, parse and return json
"""
# Get the log metadata
metadata = self.get_log_metadata(lines)

# Split lines by separator
results = []
sections = self.get_log_sections(lines)
for section in sections:
if not section.strip():
continue
section = section.split("\n")
section = [x.strip() for x in section if x.strip()]

# Parse the section. If this fails, we want to know
datum = run_parsing_function(section)
results.append(datum)

return {"data": results, "metadata": metadata, "spec": self.spec}
Loading

0 comments on commit 24980db

Please sign in to comment.