Microsecond-Scale Task Simulator

Naming and Organization

Simulation code resides in/sim/. Raw results are stored in the results directory in a directory named after the simulation name.

Simulations are named with the following convention: <hostname>_<date_time> or <hostname>_<date_time>_t<thread_number> for parallelized runs. sim_ is prepended to the name for results subdirectories.

To Run

You should be in the /sim/ directory

To run a single simulation:

python3 simulation.py <config_file> <description (optional)>

Flags: -d for debug output to standard out.

To run multiple simulations in parallel:

python3 run_sim.py <config_file_path> <options: -varycores, description>

Optional arguments:

-varycores: Vary the number of cores in each simulation rather than varying the load.

description: String to describe the simulation group. This will be written in results/meta_log

Example: python3 run_sim.py config.json -varycores "work stealing static allocations, 400ns overhead

Configuration

Configuration Parameter Dictionary

System Parameters

avg_system_load: (float) Offered load
num_queues: (int) Number of queues
num_threads: (int) Number of threads
mapping: (list[int]) Mapping from threads to queues. List of queue numbers indexed by thread number (i.e., mapping[thread #] = queue_number)
load_thread_count: (int) Number of cores to base the offered load on (allows actual cores and load to differ)
num_tasks: (int) Number of tasks in the simulation. Ends after this number of tasks completed or duration expires.
sim_duration: (int) Total simulated time in nanoseconds. Simulation ends after this many nanoseconds of simulated time or when the number of tasks runs out.
locking_enabled: (bool) Whether queue locks must be acquired or not
work_stealing_enabled: (bool) Enables work stealing
parking_enabled: (bool) Whether cores are allowed to park or not
fast_forward_enabled: (bool) Whether the simulation can skip time increments in which nothing will happen (not clear why this should ever be false)
record_alocations: (bool) Record time of allocations/allocation decisions and write to reallocation_record
reallocation_record: (string) File path to write recorded allocation schedule
record_steals: (bool) Record time of steals
buffer_cores_enabled: (bool) Enables reallocation policy in which a buffer of extra cores is maintained. Only one reallocation policy may be enabled.
delay_range_enabled: (bool) Enables reallocation policy in which cores are added/removed to maintain the range of average queueing delay. Only one reallocation policy may be enabled.
ws_self_checks: (bool) Whether cores check their own cores during the work stealing process.
allocation_delay: (bool) Whether there is a delay to allocating cores.
two_choices: (bool) Enables two choices when finding cores to work steal from.
oracle_enabled: (bool) Enables work stealing oracle which can find the best core to steal from immediately.
delay_flagging_enabled: (bool) Enables load balancing policy in which loaded cores flag other cores for assistance.
enqueue_choice: (bool) Enables load balancing policy in which arriving tasks are placed onto the least loaded core of available options.
random_work_steal_search: (bool) If enabled, work steal checks occur in random order.
regular_arrivals: (bool) If enabled, tasks arrive at regular intervals rather than as a Poisson process.
constant_service_time: (bool) If enabled, tasks have a constant service time rather than exponentially distributed service times.
ws_sibling_first: (bool) If enabled, cores check their sibling core first when work stealing.
enqueue_by_st_sum: (bool) If enabled, tasks are enqueued onto the queue with the lowest sum of service times.
always_check_realloc: (bool) If enabled, the system checks if it should reallocate cores constantly rather than at fixed intervals.
ideal_flag_steal: (bool) If enabled, flagging cores are able to flag the least-loaded core and stolen tasks are sorted by arrival time.
delay_range_by_service_time: (bool) If enabled, delay range is enforced by the sum of service times rather than queueing delay.
ideal_reallocation_enabled: (bool) Enables the reallocation policy in which the number of cores is set to match the number of tasks in the system. Only one reallocation policy may be enabled.
fred_reallocation: (bool) Enables the reallocation policy in which cores are allocated per-task. Requires always_check_realloc to be enabled. Only one reallocation policy may be enabled.
spin_parking_enabled: (bool) If enabled, cores may park after spinning to fill out the required work stealing time.
utilization_range_enabled: (bool) Enables the reallocation policy in which a range of CPU utilization is maintained. Only one reallocation policy may be enabled.
bimodal_service_time: (bool) If enabled, tasks are generated according to a bimodal distribution rather than an exponential distribution.
join_bounded_shortest_queue: (bool) Enables the load balancing policy in which tasks join a central queue and individual queues pull to maintain a certain length.
record_queue_lens: (bool) If enabled, record queue lengths at each reallocation decision.

Constants

AVERAGE_SERVICE_TIME: (int) Average service time of tasks in ns.
WORK_STEAL_CHECK_TIME: (int) Time to check if a core can be stolen from in ns.
WORK_STEAL_TIME: (int) Time to steal from another core in ns.
MINIMUM_WORK_SEARCH_TIME: (int) Minimum time that a core must work steal before parking in ns.
LOCAL_QUEUE_CHECK_TIME: (int) Time to check local queue for tasks in ns.
ALLOCATION_TIME: (int) Time required to allocate a core in ns.
BUFFER_CORE_COUNT_MIN: (int) Minimum number of buffer cores allowed. If both count and percent are defined, takes the higher value.
BUFFER_CORE_COUNT_MAX: (int) Maximum number of buffer cores allowed. If both count and percent are defined, takes the count value.
BUFFER_CORE_PCT_MIN: (int) Minimum percent of cores that should be buffer cores. If both count and percent are defined, takes the higher value.
BUFFER_CORE_PCT_MAX: (int) Maximum percent of cores that should be buffer cores. If both count and percent are defined, takes the count value.
REALLOCATION_THRESHOLD_MIN: (int) Minimum allowed queueing delay for delay range in ns.
REALLOCATION_THRESHOLD_MAX: (int) Maximum allowed queueing delay for delay range in ns.
DELAY_THRESHOLD: (int) Queueing delay threshold to trigger flagging in ns.
FLAG_STEAL_DELAY: (int) Overhead for flag stealing in ns.
ENQUEUE_CHOICES: (int) Number of enqueue choices.
ENQUEUE_PENALTY: (int) Overhead for determining enqueue location in ns.
ALLOCATION_PAUSE: (int) Time before another allocation decision can be made if a core is added in ns.
ALLOCATION_THRESHOLD: (int) Threshold for queueing delay which causes a core to be added in default core allocation policy in ns.
REQUEUE_PENALTY: (int) Overhead for moving a task between queues in ns. Applies to per-task movement of tasks.
UTILIZATION_MIN: (int) Minimum allowable CPU utilization in utilization range.
UTILIZATION_MAX: (int) Maximum allowable CPU utilization in utilization range.
FLAG_OPTIONS: (int) Number of options when flagging.
WORK_STEAL_CHOICES: (int) Number of choices when work stealing.
QUEUE_BOUND: (int) Bounded size of queue for JBSQ.

To Analyze Results

python3 analysis.py <simulations> <output_file> <ignored_time>

Arguments:

simulations: Specifies simulations to analyze. Can be (1) a file with a list of simulation names, (2) the name of a single run, or (3) the name of a run with multiple parallel threads.
output_file: Output file for results
ignored_time: Percentage of simulation time (as an int) to drop from the beginning of the data.

To Delete Old Results

Removes output files associated with the simulation and removes the line in the meta log. python3 del_old_results.py <run_name>

Flags:

-delconf: Delete the saved version of the configuration file as well.

Deletes configuration record for the given simulation and deletes the associated line in the meta log. python3 del_config_record.py <run_name>

minseok1335 / scheduling-policies-sim