geopm  3.1.1.dev272+gdfb40a8d
GEOPM - Global Extensible Open Power Manager
Classes | Public Types | Public Member Functions | Static Public Member Functions | Protected Attributes | Static Protected Attributes | List of all members
geopm::PowerBalancerAgent Class Reference

#include <PowerBalancerAgent.hpp>

Inheritance diagram for geopm::PowerBalancerAgent:
Inheritance graph
[legend]
Collaboration diagram for geopm::PowerBalancerAgent:
Collaboration graph
[legend]

Classes

class  LeafRole
 
class  MeasureRuntimeStep
 
class  ReduceLimitStep
 
class  Role
 
class  RootRole
 
class  SendDownLimitStep
 
class  Step
 
class  TreeRole
 

Public Types

enum  m_policy_e {
  M_POLICY_CPU_POWER_LIMIT , M_POLICY_STEP_COUNT , M_POLICY_MAX_EPOCH_RUNTIME , M_POLICY_POWER_SLACK ,
  M_NUM_POLICY
}
 
enum  m_sample_e {
  M_SAMPLE_STEP_COUNT , M_SAMPLE_MAX_EPOCH_RUNTIME , M_SAMPLE_SUM_POWER_SLACK , M_SAMPLE_MIN_POWER_HEADROOM ,
  M_NUM_SAMPLE
}
 
enum  m_trace_sample_e {
  M_TRACE_SAMPLE_POLICY_CPU_POWER_LIMIT , M_TRACE_SAMPLE_POLICY_STEP_COUNT , M_TRACE_SAMPLE_POLICY_MAX_EPOCH_RUNTIME , M_TRACE_SAMPLE_POLICY_POWER_SLACK ,
  M_TRACE_SAMPLE_ENFORCED_POWER_LIMIT , M_TRACE_NUM_SAMPLE
}
 
enum  m_step_e { M_STEP_SEND_DOWN_LIMIT = 0L , M_STEP_MEASURE_RUNTIME , M_STEP_REDUCE_LIMIT , M_NUM_STEP }
 

Public Member Functions

 PowerBalancerAgent (PlatformIO &platform_io, const PlatformTopo &platform_topo, std::shared_ptr< SampleAggregator > sample_agg, std::vector< std::shared_ptr< PowerBalancer > > power_balancer, double min_power, double max_power, std::shared_ptr< Waiter > waiter)
 
 PowerBalancerAgent ()
 
virtual ~PowerBalancerAgent ()=default
 
void init (int level, const std::vector< int > &fan_in, bool is_level_root) override
 Set the level where this Agent is active and push signals/controls for that level. More...
 
void validate_policy (std::vector< double > &policy) const override
 Called by Controller to validate incoming policy values and configure defaults requested in incoming policy. Policy sender can request default value with 'NaN' at the desired offset in the policy vector. Returned policy should not contain 'NaN's and be consumeable by descend and adjust_platform. More...
 
void split_policy (const std::vector< double > &in_policy, std::vector< std::vector< double > > &out_policy) override
 Called by Controller to split policy for children at next level down the tree. More...
 
bool do_send_policy (void) const override
 Called by Controller to determine if new policy values should be sent down the tree to the Agent's children. More...
 
void aggregate_sample (const std::vector< std::vector< double > > &in_sample, std::vector< double > &out_sample) override
 Aggregate samples from children for the next level up the tree. More...
 
bool do_send_sample (void) const override
 
void adjust_platform (const std::vector< double > &in_policy) override
 Adjust the platform settings based the policy from above. More...
 
bool do_write_batch (void) const override
 Called by the Controller to decide whether to call write_batch() to update platform controls. More...
 
void sample_platform (std::vector< double > &out_sample) override
 Read signals from the platform and interpret/aggregate these signals to create a sample which can be sent up the tree. More...
 
void wait (void) override
 Called by Controller to wait for sample period to elapse. This controls the cadence of the Controller main loop. More...
 
std::vector< std::pair< std::string, std::string > > report_header (void) const override
 Custom fields that will be added to the report header when this agent is used. More...
 
std::vector< std::pair< std::string, std::string > > report_host (void) const override
 Custom fields for the host section of the report. More...
 
std::map< uint64_t, std::vector< std::pair< std::string, std::string > > > report_region (void) const override
 Custom fields for each region in the report. More...
 
std::vector< std::string > trace_names (void) const override
 Column headers to be added to the trace. More...
 
std::vector< std::function< std::string(double)> > trace_formats (void) const override
 Returns format string for each column added to the trace. More...
 
void trace_values (std::vector< double > &values) override
 Called by Controller to get latest values to be added to the trace. More...
 
void enforce_policy (const std::vector< double > &policy) const override
 Enforce the policy one time with PlatformIO::write_control(). Called to enforce static policies in the absence of a Controller. More...
 
- Public Member Functions inherited from geopm::Agent
 Agent ()=default
 
virtual ~Agent ()=default
 

Static Public Member Functions

static std::string plugin_name (void)
 
static std::unique_ptr< Agentmake_plugin (void)
 
static std::vector< std::string > policy_names (void)
 
static std::vector< std::string > sample_names (void)
 
static std::string format_step_count (double step)
 
- Static Public Member Functions inherited from geopm::Agent
static std::vector< std::string > agent_names (void)
 
static std::unique_ptr< Agentmake_unique (const std::string &agent_name)
 
static int num_policy (const std::map< std::string, std::string > &dictionary)
 Used to look up the number of values in the policy vector sent down the tree for a specific Agent. This should be called with the dictionary returned by agent_factory().dictionary(agent_name) for the Agent of interest. More...
 
static int num_policy (const std::string &agent_name)
 Used to look up the number of values in the policy vector sent down the tree for a specific Agent. More...
 
static int num_sample (const std::map< std::string, std::string > &dictionary)
 Used to look up the number of values in the sample vector sent up the tree for a specific Agent. This should be called with the dictionary returned by agent_factory().dictionary(agent_name) for the Agent of interest. More...
 
static int num_sample (const std::string &agent_name)
 Used to look up the number of values in the sample vector sent up the tree for a specific Agent. More...
 
static std::vector< std::string > policy_names (const std::map< std::string, std::string > &dictionary)
 Used to look up the names of values in the policy vector sent down the tree for a specific Agent. This should be called with the dictionary returned by agent_factory().dictionary(agent_name) for the Agent of interest. More...
 
static std::vector< std::string > policy_names (const std::string &agent_name)
 Used to look up the names of values in the policy vector sent down the tree for a specific Agent. More...
 
static std::vector< std::string > sample_names (const std::map< std::string, std::string > &dictionary)
 Used to look up the names of values in the sample vector sent up the tree for a specific Agent. This should be called with the dictionary returned by agent_factory().dictionary(agent_name) for the Agent of interest. More...
 
static std::vector< std::string > sample_names (const std::string &agent_name)
 Used to look up the names of values in the sample vector sent up the tree for a specific Agent. More...
 
static std::map< std::string, std::string > make_dictionary (const std::vector< std::string > &policy_names, const std::vector< std::string > &sample_names)
 Used to create a correctly-formatted dictionary for an Agent at the time the Agent is registered with the factory. Concrete Agent classes may provide policy_names() and sample_names() methods to provide the vectors to be passed to this method. More...
 
static void aggregate_sample (const std::vector< std::vector< double > > &in_sample, const std::vector< std::function< double(const std::vector< double > &)> > &agg_func, std::vector< double > &out_sample)
 Generically aggregate a vector of samples given a vector of aggregation functions. This helper method applies a different aggregation function to each sample element while aggregating across child samples. More...
 

Protected Attributes

PlatformIO & m_platform_io
 
const PlatformTopo & m_platform_topo
 
std::shared_ptr< SampleAggregatorm_sample_agg
 
std::shared_ptr< Rolem_role
 
std::vector< std::shared_ptr< PowerBalancer > > m_power_balancer
 
double m_power_tdp
 
bool m_do_send_sample
 
bool m_do_send_policy
 
bool m_do_write_batch
 
const double M_MIN_PKG_POWER_SETTING
 
const double M_MAX_PKG_POWER_SETTING
 
const double M_TIME_WINDOW
 
std::shared_ptr< Waiterm_waiter
 

Static Protected Attributes

static constexpr double M_WAIT_SEC = 0.005
 

Additional Inherited Members

- Static Public Attributes inherited from geopm::Agent
static const std::string M_PLUGIN_PREFIX = "libgeopmagent_"
 

Member Enumeration Documentation

◆ m_policy_e

Enumerator
M_POLICY_CPU_POWER_LIMIT 

The power cap enforced on average over all nodes running the application. This has value 0.0 except in two cases. In the first case this is the M_SEND_DOWN_LIMIT step at the beginning of the application run. This value will also be non-zero in the case where the resource manager has requested a new budget for the application, and thus, the algorithm must be restarted at step M_SEND_DOWN_LIMIT.

M_POLICY_STEP_COUNT 

Step that the root is providing a policy for. The parent has received a sample matching this step in the last walk up the tree, except in the case where the root Agent has recently been updated with a new policy; in this case the step will be M_SEND_DOWN_LIMIT and the policy indexed by M_POLICY_POWER_CAP will have a non-zero value.

M_POLICY_MAX_EPOCH_RUNTIME 

The largest runtime reported by any leaf agent since the last redistribution of power. This will have value 0.0 until all leaf agents have reported a runtime to the root agent.

M_POLICY_POWER_SLACK 

This value is updated in step M_STEP_ADJUST_LIMIT to the amount that each leaf agent should increase their power limit by calling: power_cap(current_limit + slack) by before starting the algorithm over again at step M_STEP_MEASURE_RUNTIME. For all other steps this value is 0.0.

M_NUM_POLICY 

Number of steps in each iteration of the balancing algorithm.

◆ m_sample_e

Enumerator
M_SAMPLE_STEP_COUNT 

The the step counter that is currently in execution. Note that the step is equal to the step counter modulo M_NUM_STEP and is reset each time a new power cap is provided by sending a policy with a non-zero M_POLICY_POWER_CAP field.

M_SAMPLE_MAX_EPOCH_RUNTIME 

Maximum expected runtime for any node below.

M_SAMPLE_SUM_POWER_SLACK 

The sum of all slack power available from children below the agent.

M_SAMPLE_MIN_POWER_HEADROOM 

Smallest difference between maximum power limit and current power limit for any node below.

M_NUM_SAMPLE 

Number of elements in a sample vector.

◆ m_step_e

Enumerator
M_STEP_SEND_DOWN_LIMIT 

On first iteration send down resource manager average limit requested, otherwise send down average excess power.

M_STEP_MEASURE_RUNTIME 

Measure epoch runtime several times and apply median filter. Aggregate epoch runtime up tree by applying maximum filter to measured values. Propagate down from root the longest recorded runtime from any node.

M_STEP_REDUCE_LIMIT 

Decrease power limit on all nodes (other than the slowest) until epoch runtime matches the slowest. Aggregate amount power limit was reduced in last step up the tree with sum filter. (Go to M_STEP_SEND_DOWN_LIMIT next).

M_NUM_STEP 

Number of steps in process.

◆ m_trace_sample_e

Enumerator
M_TRACE_SAMPLE_POLICY_CPU_POWER_LIMIT 
M_TRACE_SAMPLE_POLICY_STEP_COUNT 
M_TRACE_SAMPLE_POLICY_MAX_EPOCH_RUNTIME 
M_TRACE_SAMPLE_POLICY_POWER_SLACK 
M_TRACE_SAMPLE_ENFORCED_POWER_LIMIT 
M_TRACE_NUM_SAMPLE 

Constructor & Destructor Documentation

◆ PowerBalancerAgent() [1/2]

geopm::PowerBalancerAgent::PowerBalancerAgent ( PlatformIO &  platform_io,
const PlatformTopo &  platform_topo,
std::shared_ptr< SampleAggregator sample_agg,
std::vector< std::shared_ptr< PowerBalancer > >  power_balancer,
double  min_power,
double  max_power,
std::shared_ptr< Waiter waiter 
)

◆ PowerBalancerAgent() [2/2]

geopm::PowerBalancerAgent::PowerBalancerAgent ( )

◆ ~PowerBalancerAgent()

virtual geopm::PowerBalancerAgent::~PowerBalancerAgent ( )
virtualdefault

Member Function Documentation

◆ adjust_platform()

void geopm::PowerBalancerAgent::adjust_platform ( const std::vector< double > &  in_policy)
overridevirtual

Adjust the platform settings based the policy from above.

Parameters
[in]policySettings for each control in the policy.

Implements geopm::Agent.

◆ aggregate_sample()

void geopm::PowerBalancerAgent::aggregate_sample ( const std::vector< std::vector< double > > &  in_sample,
std::vector< double > &  out_sample 
)
overridevirtual

Aggregate samples from children for the next level up the tree.

Parameters
[in]in_sampleVector of sample vectors, one from each child.
[out]out_sampleAggregated sample values to be sent up to the parent.

Implements geopm::Agent.

◆ do_send_policy()

bool geopm::PowerBalancerAgent::do_send_policy ( void  ) const
overridevirtual

Called by Controller to determine if new policy values should be sent down the tree to the Agent's children.

Returns
True if the policy has been updated since last call.

Implements geopm::Agent.

◆ do_send_sample()

bool geopm::PowerBalancerAgent::do_send_sample ( void  ) const
overridevirtual
Returns
True if new samples were read since last call.

Implements geopm::Agent.

◆ do_write_batch()

bool geopm::PowerBalancerAgent::do_write_batch ( void  ) const
overridevirtual

Called by the Controller to decide whether to call write_batch() to update platform controls.

Returns
True if platform was adjusted, false otherwise.

Implements geopm::Agent.

◆ enforce_policy()

void geopm::PowerBalancerAgent::enforce_policy ( const std::vector< double > &  policy) const
overridevirtual

Enforce the policy one time with PlatformIO::write_control(). Called to enforce static policies in the absence of a Controller.

Implements geopm::Agent.

◆ format_step_count()

std::string geopm::PowerBalancerAgent::format_step_count ( double  step)
static

◆ init()

void geopm::PowerBalancerAgent::init ( int  level,
const std::vector< int > &  fan_in,
bool  is_level_root 
)
overridevirtual

Set the level where this Agent is active and push signals/controls for that level.

Parameters
[in]levelLevel of the tree where this agent is active. Note that only agents at level zero execute sample_platform() and adjust_platform().
[in]fan_inVector over level giving the the number of Agents that report to each root Agent operating at the level.
[in]is_level_rootTrue if the agent plays the role of root of the level. Only root agents for a level execute ascend() and descend().

Implements geopm::Agent.

◆ make_plugin()

std::unique_ptr< Agent > geopm::PowerBalancerAgent::make_plugin ( void  )
static

◆ plugin_name()

std::string geopm::PowerBalancerAgent::plugin_name ( void  )
static

◆ policy_names()

std::vector< std::string > geopm::PowerBalancerAgent::policy_names ( void  )
static

◆ report_header()

std::vector< std::pair< std::string, std::string > > geopm::PowerBalancerAgent::report_header ( void  ) const
overridevirtual

Custom fields that will be added to the report header when this agent is used.

Implements geopm::Agent.

◆ report_host()

std::vector< std::pair< std::string, std::string > > geopm::PowerBalancerAgent::report_host ( void  ) const
overridevirtual

Custom fields for the host section of the report.

Implements geopm::Agent.

◆ report_region()

std::map< uint64_t, std::vector< std::pair< std::string, std::string > > > geopm::PowerBalancerAgent::report_region ( void  ) const
overridevirtual

Custom fields for each region in the report.

Implements geopm::Agent.

◆ sample_names()

std::vector< std::string > geopm::PowerBalancerAgent::sample_names ( void  )
static

◆ sample_platform()

void geopm::PowerBalancerAgent::sample_platform ( std::vector< double > &  out_sample)
overridevirtual

Read signals from the platform and interpret/aggregate these signals to create a sample which can be sent up the tree.

Parameters
[out]sampleVector of agent specific sample values to be sent up the tree.

Implements geopm::Agent.

◆ split_policy()

void geopm::PowerBalancerAgent::split_policy ( const std::vector< double > &  in_policy,
std::vector< std::vector< double > > &  out_policy 
)
overridevirtual

Called by Controller to split policy for children at next level down the tree.

Parameters
[in]in_policyPolicy values from the parent.
[out]out_policyVector of policies to be sent to each child.

Implements geopm::Agent.

◆ trace_formats()

std::vector< std::function< std::string(double)> > geopm::PowerBalancerAgent::trace_formats ( void  ) const
overridevirtual

Returns format string for each column added to the trace.

Reimplemented from geopm::Agent.

◆ trace_names()

std::vector< std::string > geopm::PowerBalancerAgent::trace_names ( void  ) const
overridevirtual

Column headers to be added to the trace.

Implements geopm::Agent.

◆ trace_values()

void geopm::PowerBalancerAgent::trace_values ( std::vector< double > &  values)
overridevirtual

Called by Controller to get latest values to be added to the trace.

Implements geopm::Agent.

◆ validate_policy()

void geopm::PowerBalancerAgent::validate_policy ( std::vector< double > &  policy) const
overridevirtual

Called by Controller to validate incoming policy values and configure defaults requested in incoming policy. Policy sender can request default value with 'NaN' at the desired offset in the policy vector. Returned policy should not contain 'NaN's and be consumeable by descend and adjust_platform.

Parameters
[in,out]policyPolicy replace NaN's with defaults. If a value of policy is not NaN but the value is not supported by the Agent the method will throw a geopm::Exception with error code GEOPM_ERROR_INVALID.

Implements geopm::Agent.

◆ wait()

void geopm::PowerBalancerAgent::wait ( void  )
overridevirtual

Called by Controller to wait for sample period to elapse. This controls the cadence of the Controller main loop.

Implements geopm::Agent.

Member Data Documentation

◆ m_do_send_policy

bool geopm::PowerBalancerAgent::m_do_send_policy
protected

◆ m_do_send_sample

bool geopm::PowerBalancerAgent::m_do_send_sample
protected

◆ m_do_write_batch

bool geopm::PowerBalancerAgent::m_do_write_batch
protected

◆ M_MAX_PKG_POWER_SETTING

const double geopm::PowerBalancerAgent::M_MAX_PKG_POWER_SETTING
protected

◆ M_MIN_PKG_POWER_SETTING

const double geopm::PowerBalancerAgent::M_MIN_PKG_POWER_SETTING
protected

◆ m_platform_io

PlatformIO& geopm::PowerBalancerAgent::m_platform_io
protected

◆ m_platform_topo

const PlatformTopo& geopm::PowerBalancerAgent::m_platform_topo
protected

◆ m_power_balancer

std::vector<std::shared_ptr<PowerBalancer> > geopm::PowerBalancerAgent::m_power_balancer
protected

◆ m_power_tdp

double geopm::PowerBalancerAgent::m_power_tdp
protected

◆ m_role

std::shared_ptr<Role> geopm::PowerBalancerAgent::m_role
protected

◆ m_sample_agg

std::shared_ptr<SampleAggregator> geopm::PowerBalancerAgent::m_sample_agg
protected

◆ M_TIME_WINDOW

const double geopm::PowerBalancerAgent::M_TIME_WINDOW
protected

◆ M_WAIT_SEC

constexpr double geopm::PowerBalancerAgent::M_WAIT_SEC = 0.005
staticconstexprprotected

◆ m_waiter

std::shared_ptr<Waiter> geopm::PowerBalancerAgent::m_waiter
protected

The documentation for this class was generated from the following files: