Built-in Policies

DMR ships three built-in policies. Each one is obtained from a constructor that returns a ready-to-use DMRPolicy *, which you hand to dmr_set_policy.

Constructor	Header	Description
dmr_get_policy_round()	dmr_test_policies.h	Multiplies the current node count by a stride, wrapping back to the minimum
dmr_get_policy_list()	dmr_test_policies.h	Cycles through a fixed list of node counts
dmr_get_policy_ce()	dmr_CE_policy.h	Targets a communication efficiency value, using TALP metrics

dmr.h does not pull these headers in: include the one you need explicitly. Each policy is also compiled in only if its CMake flag is on, DMR_BUILD_TEST_POLICIES for round and list, DMR_BUILD_CE_POLICY for CE. Both are on by default, the second one only in TALP builds. The declarations, guards, and flags are in Policy Headers.

Round policy

Multiplies the current node count by stride. When the result would exceed the maximum, it wraps back to the minimum.

#include "dmr.h"
#include "dmr_test_policies.h"

dmr_set_policy(dmr_get_policy_round());

while (should_keep_running()) {
    DMR_AUTO(dmr_check(USE_POLICY), save(), (void)NULL, cleanup());
    do_work();
}

With MIN=1, MAX=8, STRIDE=2 the sequence is 1 → 2 → 4 → 8 → 1 → …

Parameter	Environment variable	Default
Minimum nodes	DMR_DEFAULT_POLICY_MIN	1
Maximum nodes	DMR_DEFAULT_POLICY_MAX	1
Stride (multiplier)	DMR_DEFAULT_POLICY_STRIDE	2

The defaults are also the compile-time fallbacks, settable as CMake variables of the same name.

List policy

Cycles through the hardcoded sequence {2, 3, 4, 5, 6, 7, 8, 7, 6, 5, 4, 3, 2, 1}, indexed by dmr_get_reconfig_count(). Intended for testing and benchmarking.

#include "dmr.h"
#include "dmr_test_policies.h"

dmr_set_policy(dmr_get_policy_list());

No parameters, and no minimum or maximum: the list itself is the bound.

CE policy

Reads the accumulated communication efficiency from TALP in populate, then in run adjusts the node count to push that efficiency towards a target:

change      = round((target_ce - current_ce) * -sensitivity)
target_nodes = clamp(current_nodes + change, min_nodes, max_nodes)

Efficiency below the target means the job is spending too much time in communication, so change is negative and the policy shrinks; efficiency above the target lets it grow. If TALP reports no usable metrics for this iteration, the policy stays.

#include "dmr.h"
#include "dmr_CE_policy.h"

dmr_set_policy(dmr_get_policy_ce());

Parameter	Environment variable	Default
Minimum nodes	DMR_DEFAULT_POLICY_MIN	1
Maximum nodes	DMR_DEFAULT_POLICY_MAX	1
Target CE	DMR_TALP_TARGET_CE	0.8
Sensitivity	DMR_TALP_SENSITIVITY	15

Requires a build with DMR_USE_TALP=ON and the job launched so that TALP is collecting metrics; see Configuration.

Changing parameters at runtime

Every constructor reads its environment variables each time it is called, and returns the same static object, so re-registering the policy is how you re-parameterise it:

setenv("DMR_DEFAULT_POLICY_MAX", "16", 1);
dmr_set_policy(dmr_get_policy_round());   /* collective: all ranks */

All ranks must call dmr_set_policy, and they must see the same values, otherwise ranks will disagree on the suggestion. For anything more elaborate, write a custom policy.