sys Namespace Reference

Utilities for hardware and system interop. More...

Enumerations
enum class	PrefetchTo : int {   L1Cache = 3 , L2Cache = 2 , L3Cache = 1 , NonTemporal = 0 ,   WriteL1Cache = 7 , WriteL2Cache = 6 }
	The cache location to prefetch data to. More...

Functions
void	abort (const std::string &message)
	Abort the program with an error message. More...
void	attach_debugger ()
	Provide an infinite loop to attach a debugger during startup. Useful for debugging MPI runs. More...
void	exit (int exit_code=0)
	Exit the program normally. This should only be called once over all processors.
int	number_of_procs ()
	Number of processing elements.
int	my_proc ()
	Index of my processing element.
int	number_of_nodes ()
	Number of nodes.
int	my_node ()
	Index of my node.
int	procs_on_node (int node_index)
	Number of processing elements on the given node.
int	my_local_rank ()
	The local index of my processing element on my node. This is in the interval 0, ..., procs_on_node(my_node()) - 1.
int	first_proc_on_node (int node_index)
	Index of first processing element on the given node.
int	node_of (int proc_index)
	Index of the node for the given processing element.
int	local_rank_of (int proc_index)
	The local index for the given processing element on its node.
double	wall_time ()
	The elapsed wall time in seconds.
template<PrefetchTo CacheLocation>
void	prefetch (const void *address_to_prefetch)
	Prefetch data into a specific level of data cache.
std::ostream &	operator<< (std::ostream &os, PrefetchTo cache_location)
std::string	pretty_wall_time (double total_seconds)
	Format the wall time in DD-HH:MM:SS format. More...
std::string	pretty_wall_time ()
	Format the wall time in DD-HH:MM:SS format. More...

Detailed Description

Utilities for hardware and system interop.

Enumeration Type Documentation

PrefetchTo

enum class sys::PrefetchTo : int

strong

The cache location to prefetch data to.

Enumerator
L1Cache	Prefetch into L1 data cache. Typically the fastest CPU cache, about 32 KB in size.
L2Cache	Prefetch into L2 data cache. Typically the second fastest CPU cache, about 128 or 256 KB in size.
L3Cache	Prefetch into L3 data cache. Typically the slowest CPU cache and is shared among multiple cores with sizes varying by factors of several.
NonTemporal	Non-temporal is an element that is unlikely to be re-used. E.g., read once, written but never read.
WriteL1Cache	Prefetch to L1 data cache for writing.
WriteL2Cache	Prefetch to L2 data cache for writing.

Function Documentation

attach_debugger()

void sys::attach_debugger

(

)

Provide an infinite loop to attach a debugger during startup. Useful for debugging MPI runs.

Each MPI rank writes a file name spectre_pid_#_host_NAME to the working directory. This allows you to attach GDB to the running process using gdb --pid=PID, once for each MPI rank. You must then halt the program using C-c and then call set var i = 7 inside GDB. Once you've done this on each MPI rank, you can have each MPI rank continue.

To add support for attaching to a debugger in an executable, you must add sys::attach_debugger to the Parallel::charmxx::register_init_node_and_proc init node functions. Then, when you launch the executable launch it as

SPECTRE_ATTACH_DEBUGGER=1 mpirun -np N ...

The environment variable SPECTRE_ATTACH_DEBUGGER being set tells the code to allow attaching from a debugger.