232 lines
11 KiB
TOML
232 lines
11 KiB
TOML
# This file is part of Nanos6 and is licensed under the terms contained in the COPYING file
|
|
#
|
|
# Copyright (C) 2020-2021 Barcelona Supercomputing Center (BSC)
|
|
|
|
# The Nanos6 default configuration file. More details about the configuration options and the possible
|
|
# values are available on the OmpSs-2 User Guide (https://pm.bsc.es/ftp/ompss-2/doc/user-guide/) or on
|
|
# the README.md in the Nanos6 repository (https://github.com/bsc-pm/nanos6)
|
|
|
|
[version]
|
|
# Choose whether the runtime runs with debug symbols and internal asserts. Enabling this option
|
|
# may produce significant overheads, so production or performance executions should disable this
|
|
# option. Default is false
|
|
debug = false
|
|
# Choose the dependency system implementation. Default is "discrete"
|
|
# Possible values: "discrete", "regions"
|
|
dependencies = "discrete"
|
|
# Choose the instrumentation variant to run. Default is "none"
|
|
# Possible values: "none", "ctf", "extrae", "graph", "lint", "stats", "verbose"
|
|
instrument = "ovni"
|
|
|
|
[turbo]
|
|
# Choose whether enabling floating-point unit optimizations in all Nanos6 threads. Enabling this
|
|
# option can speedup the floating-point computations in user tasks, but may produce imprecise
|
|
# results. It enables flush-to-zero (FZ) and denormals are zero (DAZ) optimizations in Intel®
|
|
# processors. Default is false
|
|
enabled = false
|
|
|
|
[scheduler]
|
|
# Choose the task scheduling policy. Default is "fifo"
|
|
# Possible values: "fifo", "lifo"
|
|
policy = "fifo"
|
|
# Enable the immediate successor feature to improve cache data reutilization between successor
|
|
# tasks. If enabled, when a CPU finishes a task it starts executing the successor task (computed
|
|
# through their data dependencies). Default is true
|
|
immediate_successor = true
|
|
# Indicate whether the scheduler should consider task priorities defined by the user in the
|
|
# task's priority clause. Default is true
|
|
priority = true
|
|
|
|
[cpumanager]
|
|
# The underlying policy of the CPU manager for the handling of CPUs. Default is "default", which
|
|
# corresponds to "hybrid"
|
|
# Possible values: "default", "idle", "busy", "hybrid", "lewi", "greedy"
|
|
policy = "default"
|
|
# The maximum number of iterations to busy wait for before idling. Default is "240000". Only
|
|
# works for the 'hybrid' policy. This number will be divided by the number of active CPUs to
|
|
# obtain a "busy_iters per CPU" metric for each individual CPU to busy-wait for
|
|
busy_iters = 240000
|
|
|
|
[taskfor]
|
|
# Choose the total number of CPU groups that will execute the worksharing tasks (taskfors). Default
|
|
# is none (not set), which means that the runtime will create one taskfor group per NUMA node
|
|
# groups = 1
|
|
# Indicate whether should print the taskfor groups information
|
|
report = false
|
|
|
|
[throttle]
|
|
# Enable throttle to stop creating tasks when certain conditions are met. Default is false
|
|
enabled = false
|
|
# Maximum number of child tasks that can be created before throttling. Default is 5000000
|
|
tasks = 5000000
|
|
# Maximum memory pressure (percent of max_memory) before throttling. Default is 70 (%)
|
|
pressure = 70 # %
|
|
# Maximum memory that can be used by the runtime. Default is "0", which equals half of system memory
|
|
max_memory = "0"
|
|
# Evaluation interval (us). Each time this amount of time is elapsed, the throttle system queries
|
|
# the memory allocator statistics and evaluates the current memory pressure. A higher interval
|
|
# results in less accurate pressure estimation, but a lower interval introduces noticeable overhead,
|
|
# as reading memory statistics requires a global lock on the allocator to aggregate per-thread
|
|
# statistics. Default is 1000
|
|
polling_period_us = 1000
|
|
|
|
[numa]
|
|
# Enable NUMA tracking of task data. NUMA tracking consists of annotating the NUMA location
|
|
# of data to be later scheduled based on this information. When using "auto" this feature is
|
|
# enabled in the first allocation done using the Nanos6 NUMA API. Default is "auto"
|
|
# Possible values: "auto", "on", "off"
|
|
tracking = "auto"
|
|
# Indicate whether should print the NUMA bitmask of each NUMA wildcards
|
|
report = false
|
|
# Use the page auto-discovery mechanism to detect the system's page size
|
|
# Default is true, which is useful in systems with THP enabled
|
|
# Set to false will use the default page size, which is arch-dependent
|
|
discover_pagesize = true
|
|
|
|
[hardware_counters]
|
|
# Enable the verbose mode of hardware counters, printing a small summary of metrics at the
|
|
# end of the execution. Default is false
|
|
verbose = false
|
|
# The verbose file's name. Default is "nanos6-output-hwcounters.txt"
|
|
verbose_file = "nanos6-output-hwcounters.txt"
|
|
[hardware_counters.papi]
|
|
# Enable the PAPI backend of the hardware counters module. Default is false
|
|
enabled = false
|
|
# The list of PAPI counters to read. Default is "PAPI_TOT_INS" and "PAPI_TOT_CYC"
|
|
counters = [
|
|
"PAPI_TOT_INS",
|
|
"PAPI_TOT_CYC"
|
|
]
|
|
[hardware_counters.rapl]
|
|
# Enable the RAPL backend of the hardware counters module for runtime-wise energy
|
|
# metrics. Default is false
|
|
enabled = false
|
|
|
|
[monitoring]
|
|
# Indicate whether enabling the Monitoring of tasks and CPUs, which allows the usage of
|
|
# prediction-based policies. Disabled by default
|
|
enabled = false
|
|
# Indicate whether enabling the "Wisdom" mechanism of Monitoring, saving normalized metrics for
|
|
# future executions and loading previously saved metrics when the runtime initializes. Default
|
|
# is false
|
|
wisdom = false
|
|
# Enable the verbose mode of Monitoring, which prints a detailed summary of task type metrics
|
|
# at the end of the execution. Default is true
|
|
verbose = true
|
|
# The verbose file's name. Default is "output-monitoring.txt"
|
|
verbose_file = "output-monitoring.txt"
|
|
# The prediction rate at which CPU usage predictions are infered. Default is once every 100µs
|
|
cpuusage_prediction_rate = 100 # µs
|
|
# The number of samples (window) of the normalized exponential moving average for predictions
|
|
# Default is 20
|
|
rolling_window = 20
|
|
|
|
[devices]
|
|
|
|
[instrument]
|
|
[instrument.ctf]
|
|
# Choose the temporary directory where to store intermediate CTF files. Default is none
|
|
# (not set), which means that $TMPDIR will be used if present, or /tmp otherwise
|
|
# tmpdir = "/tmp"
|
|
[instrument.ctf.converter]
|
|
# Indicate whether the trace converter should automatically generate the trace after
|
|
# executing a program with CTF instrumentation. Default is true
|
|
enabled = true
|
|
# Use the fast converter. This feature is experimental and generates a trace compatible
|
|
# with just a subset of Paraver cfgs. Default is false
|
|
fast = false
|
|
# Indicate the location of the ctf2prv converter script. Default is none (not set),
|
|
# which means that the $CTF2PRV will be used if present, or ctf2prv in $PATH
|
|
# otherwise
|
|
# location = "path/to/ctf2prv"
|
|
# Choose the events that will be traced
|
|
[instrument.ctf.events]
|
|
# Linux Kernel events options. Nanos6 can collect Linux kernel internal events using the
|
|
# perf_event_open system call. This requires to set /proc/sys/kernel/perf_event_paranoid
|
|
# to -1 and read permissions for either /sys/kernel/tracing or /sys/kernel/debug/tracing.
|
|
# Events can be selected individually and/or using Nanos6-defined presets
|
|
[instrument.ctf.events.kernel]
|
|
# Select one or more predefined sets of events. Available options are:
|
|
# - preemption: Trace context switches and interrupts
|
|
# - context_switch: Trace context switches
|
|
# - syscall: Trace all kernel syscalls entry and exit
|
|
# presets = [
|
|
# "preemption"
|
|
# ]
|
|
# Provide a file with a list of individual events to enable, one event per line.
|
|
# Lines starting with "#" are omitted.
|
|
# file = "./nanos6_kernel_events"
|
|
# Exclude selected events. Disable events even if enabled in a preset or the
|
|
# user provided file.
|
|
# exclude = [
|
|
# "sys_enter_gettimeofday",
|
|
# "sys_exit_gettimeofday"
|
|
# ]
|
|
[instrument.extrae]
|
|
# Indicate whether the trace should show the activity of the threads instead of the CPUs'
|
|
# activity. Default is false, which shows the activity of CPUs
|
|
as_threads = false
|
|
# Choose the detail level of the information generated in extrae traces. Default is 1
|
|
detail_level = 1
|
|
[instrument.graph]
|
|
# Indicate whether the resulting PDF should be opened automatically with the display
|
|
# command. Default is false
|
|
display = false
|
|
# Choose the command to be used to automatically display the resulting PDF
|
|
display_command = "xdg-open"
|
|
# Indicate whether the graph nodes should remove the directories from the source code
|
|
# file names. Default is false
|
|
shorten_filenames = false
|
|
# Indicate whether should show the internal data structures that determine when tasks
|
|
# are ready. Default is false
|
|
show_dependency_structures = false
|
|
# Include the information about the range of data or region that is covered when showing
|
|
# internal data structures. Default is false
|
|
show_regions = false
|
|
# Indicate whether should show the internal data structures that do not determine dependencies
|
|
# or that are redundant by transitivity. Default is false
|
|
show_spurious_dependency_structures = false
|
|
# Choose whether should force future and previous dependencies to be shown with different
|
|
# graphical attributes. Default is false
|
|
show_dead_dependencies = false
|
|
# Include the internal data structures after they are no longer relevant. Default is false
|
|
show_dead_dependency_structures = false
|
|
# Instead of trying to collapse in one step as many related changes as possible, show one
|
|
# at a time. Default is false
|
|
show_all_steps = false
|
|
# Show superaccess links. Default is true
|
|
show_superaccess_links = true
|
|
# Choose whether to emit a table next to the graph with a description of the changes in
|
|
# each frame. Default is false
|
|
show_log = false
|
|
[instrument.stats]
|
|
# The outfile file for emitting the statistics. Default is the standard error
|
|
output_file = "/dev/stderr"
|
|
[instrument.verbose]
|
|
# Output device or file for verbose log. Default is "/dev/stderr"
|
|
output_file = "/dev/stderr"
|
|
# Print timestamps on log. Default is true
|
|
timestamps = true
|
|
# Delay verbose output to prevent mixing with application output. Default is false
|
|
dump_only_on_exit = false
|
|
# Verbose log concepts to display. Possible values on README.md
|
|
areas = ["all", "!ComputePlaceManagement", "!DependenciesByAccess", "!DependenciesByAccessLinks",
|
|
"!DependenciesByGroup", "!LeaderThread", "!TaskStatus", "!ThreadManagement"]
|
|
|
|
[memory]
|
|
|
|
[misc]
|
|
# Stack size of threads created by the runtime. Default is 8M
|
|
stack_size = "8M"
|
|
|
|
[loader]
|
|
# Enable verbose output of the loader, to debug dynamic linking problems. Default is false
|
|
verbose = false
|
|
# Choose whether to print a warning when there is any invalid NANOS6 environment variables
|
|
# defined when running a program. Default is true
|
|
warn_envars = true
|
|
# Path to the nanos6 libraries. Default is none
|
|
# library_path = "path/to/nanos6/lib"
|
|
# Prefix for the report printing all the runtime variables. Default is none
|
|
# report_prefix = "#"
|