Fastsim parallel integration

system:

  num_cdus: 6

  racks_per_cdu: 6

  nodes_per_rack: 80

  rectifiers_per_rack: 6

  chassis_per_rack: 1

  nodes_per_blade: 1

  switches_per_chassis: 5

  nics_per_node: 2

  rectifiers_per_chassis: 5

  nodes_per_rectifier: 4

  missing_racks: []

  down_nodes: []

  cpus_per_node: 1

  gpus_per_node: 4

  cpu_peak_flops: 396800000000.0

  gpu_peak_flops: 7800000000000.0

  cpu_fp_ratio: 0.69

  gpu_fp_ratio: 0.69

power:

  power_gpu_idle: 75

  power_gpu_max: 300

  power_cpu_idle: 100

  power_cpu_max: 800

  power_mem: 74.26

  power_nic: 21

  power_nvme: 45

  power_switch: 250

  power_cdu: 0

  power_update_freq: 20

  rectifier_peak_threshold: 13670

  sivoc_loss_constant: 0

  sivoc_efficiency: 1

  rectifier_loss_constant: 0

  rectifier_efficiency: 1

  power_cost: 0.094

scheduler:

  seed: 42

  job_arrival_time: 20

  mtbf: 11

  trace_quanta: 20

  min_wall_time: 3600

  max_wall_time: 43200

  ui_update_freq: 3600

  max_nodes_per_job: 3000

  job_end_probs:

    COMPLETED: 0.63

    FAILED: 0.13

    CANCELLED: 0.12

    TIMEOUT: 0.11

    NODE_FAIL: 0.01

 No newline at end of file

"""

    Load data for NREL's Kestrel cluster.

"""

import uuid

import pandas as pd

from tqdm import tqdm

from ..job import job_dict, Job

from ..utils import power_to_utilization, next_arrival

def load_data(jobs_path, **kwargs):

    """

    Reads job and job profile data from parquet files and parses them.

    Parameters

    ----------

    jobs_path : str

        The path to the jobs parquet file.

    Returns

    -------

    list

        The list of parsed jobs.

    """

    jobs_df = pd.read_parquet(jobs_path, engine='pyarrow')

    return load_data_from_df(jobs_df, **kwargs)

def load_data_from_df(jobs_df: pd.DataFrame, **kwargs):

    """

    Reads job and job profile data from parquet files and parses them.

    Requires the following fields in the DataFrame:

    - start_time (timestamp): Time execution begins (actual or expected) 	

    - job_id (int): Job ID

    - node_power_consumption (List[int]): Power consumption of the job, recorded at Node level

    - nodes_required (int): Number of nodes allocated to the job

    - cpu_power_consumption (List[int]): Power consumption of the job, recorded at CPU level (don't have this)

    - mem_power_consumption (List[int]): Power consumption of the job, recorded at Memory level (don't have this)

    - priority (int): Relative priority of the job, 0=held, 1=required nodes DOWN/DRAINED

    - job_state (string): State of the job, see enum job_states for possible values

    - wall_time (int): Actual runtime of job, in seconds

    - nodes (string): List of nodes allocated to job

    Returns

    -------

    list

        The list of parsed jobs.

    """

    config = kwargs.get('config')

    min_time = kwargs.get('min_time', None)

    reschedule = kwargs.get('reschedule')

    fastforward = kwargs.get('fastforward')

    validate = kwargs.get('validate')

    jid = kwargs.get('jid', '*')

    if fastforward: print(f"fast-forwarding {fastforward} seconds")

    # Sort jobs dataframe based on values in time_start column, adjust indices after sorting

    jobs_df = jobs_df.sort_values(by='submit_time')

    jobs_df = jobs_df[(jobs_df.start_time.between(pd.to_datetime('2024-09-01T00:00:00'), 

                                                 pd.to_datetime('2024-09-16T00:00:00'), inclusive='right') | 

                        jobs_df.end_time.between(pd.to_datetime('2024-09-01T00:00:00'), 

                                                 pd.to_datetime('2024-09-16T00:00:00'), inclusive='right') 

                                                 )].copy()

    jobs_df = jobs_df.reset_index(drop=True)

    telemetry_start_timestamp = jobs_df['start_time'].min()

    telemetry_end_timestamp = jobs_df['end_time'].max()

    telemetry_start = 0

    telemetry_end = int((telemetry_end_timestamp - telemetry_start_timestamp).total_seconds())

    # Take earliest time as baseline reference

    # We can use the start time of the first job.

    if min_time:

        time_zero = min_time

    else:

        time_zero = jobs_df['submit_time'].min()

    num_jobs = len(jobs_df)

    print("time_zero:", time_zero, "num_jobs", num_jobs)

    jobs = []

    # Map dataframe to job state. Add results to jobs list

    for jidx in tqdm(range(num_jobs - 1), total=num_jobs, desc="Processing Kestrel Jobs"):

        job_id = jobs_df.loc[jidx, 'job_id']

        account = jobs_df.loc[jidx, 'account']

        if not jid == '*': 

            if int(jid) == int(job_id): 

                print(f'Extracting {job_id} profile')

            else:

                continue

        nodes_required = jobs_df.loc[jidx, 'nodes_required']

        name = str(uuid.uuid4())[:6]

        if validate:

            cpu_power = jobs_df.loc[jidx, 'power_per_node']

            cpu_trace = cpu_power

        else:                

            cpu_power = jobs_df.loc[jidx, 'power_per_node']

            cpu_power_array = [600] if (pd.isna(cpu_power) or cpu_power == 0) else cpu_power.tolist()

            cpu_min_power = nodes_required * config['POWER_CPU_IDLE'] * config['CPUS_PER_NODE']

            cpu_max_power = nodes_required * config['POWER_CPU_MAX'] * config['CPUS_PER_NODE']

            cpu_util = power_to_utilization(cpu_power_array, cpu_min_power, cpu_max_power)

            cpu_trace = cpu_util * config['CPUS_PER_NODE']

            gpu_trace = 0

        # Priority sorting doesn't seem to be implemented at the moment

        priority = 0

        wall_time = jobs_df.loc[jidx, 'wall_time']

        end_state = jobs_df.loc[jidx, 'job_state']

        time_submit = jobs_df.loc[jidx+1, 'submit_time']

        diff = time_submit - time_zero

        if jid == '*': 

            time_offset = max(diff.total_seconds(), 0)

        else:

            # When extracting out a single job, run one iteration past the end of the job

            time_offset = config['UI_UPDATE_FREQ']

        if fastforward: time_offset -= fastforward

        if reschedule: # Let the scheduler reschedule the jobs

            scheduled_nodes = None

            time_offset = next_arrival(1/config['JOB_ARRIVAL_TIME'])

        else: # Prescribed replay

            scheduled_nodes = None

            time_offset = next_arrival(1/config['JOB_ARRIVAL_TIME'])

        trace_quanta = config['TRACE_QUANTA']

        if cpu_trace.size > 0 and time_offset >= 0:

            job_info = job_dict(nodes_required = nodes_required, 

                                name = name, 

                                account = account, 

                                cpu_trace = cpu_trace, 

                                gpu_trace = gpu_trace, 

                                ntx_trace = [], 

                                nrx_trace = [],

                                end_state = end_state, 

                                scheduled_nodes = scheduled_nodes,

                                id = job_id, 

                                priority = priority, 

                                submit_time = time_offset, 

                                time_limit = wall_time,

                                trace_quanta=trace_quanta)

            jobs.append(Job(job_info))

    return jobs, telemetry_start, telemetry_end

def node_index_to_name(index: int, config: dict):

    """ Converts an index value back to an name string based on system configuration. """

    return f"node{index:04d}"

def cdu_index_to_name(index: int, config: dict):

    return f"cdu{index:02d}"

def cdu_pos(index: int, config: dict) -> tuple[int, int]:

    """ Return (row, col) tuple for a cdu index """

    return (0, index) # TODO

 No newline at end of file

from raps.sim_config import SimConfig

from raps.system_config import SystemConfig

from bisect import bisect_right

@dataclasses.dataclass

class TickData:

            self.power_manager.set_idle(job.scheduled_nodes)

            job.current_state = JobState.COMPLETED

            job.end_time = self.current_timestep

            self.running.remove(job)

            self.jobs_completed += 1

            job_stats = job.statistics()

            self.resource_manager.free_nodes_from_job(job)

        killed_jobs = [job for job in self.running if

                       job.start_time + job.time_limit <= self.current_timestep]

                       job.end_time is not None and job.start_time + job.time_limit <= self.current_timestep]

        for job in killed_jobs:

            self.power_manager.set_idle(job.scheduled_nodes)

                          actively_considered_jobs: List,

                          all_jobs: List,

                          replay: bool,

                          autoshutdown: bool):

                          autoshutdown: bool,

                          cursor: int):

        # 1 update running time of all running jobs

        # 2 update the current_timestep of the engine (this serves as reference for most computations)

        # 3 Check if simulation should shutdown

           len(self.queue) == 0 and \

           len(self.running) == 0 and \

           not replay and \

           len(all_jobs) == 0 and \

           len(all_jobs) == cursor and \

           len(actively_considered_jobs) == 0:

            if self.debug:

                print(f"Simulaiton completed early: {self.config['system_name']} - "

            job.running_time = self.current_timestep - job.start_time

            if job.current_state != JobState.RUNNING:

                raise ValueError(f"Job is in running list, but state is not RUNNING: job.state == {job.currentstate}")

                raise ValueError(f"Job {job.id} is in running list, but state is not RUNNING: job.state == {job.current_state}")

            else:  # if job.state == JobState.RUNNING:

                # Error checks

                if job.running_time > job.time_limit:

                if job.running_time > job.time_limit and job.end_time is not None:

                    raise Exception(f"Job exceded time limit! "

                                    f"{job.running_time} > {job.time_limit}"

                                    f"\n{job}"

        # Process jobs in batches for better performance of timestep loop

        all_jobs = jobs.copy()

        submit_times = [j.submit_time for j in all_jobs]

        cursor = 0

        jobs = []

        # Batch Jobs into 6h windows based on submit_time or twice the time_delta if larger

        batch_window = max(60 * 60 * 6, 2 * time_delta)  # at least 6h

            if (self.current_timestep % batch_window == 0) or (self.current_timestep == timestep_start):

                # Add jobs that are within the batching window and remove them from all jobs

                jobs += [job for job in all_jobs if job.submit_time <= self.current_timestep + batch_window]

                all_jobs[:] = [job for job in all_jobs if job.submit_time > self.current_timestep + batch_window]

                # jobs += [job for job in all_jobs if job.submit_time <= self.current_timestep + batch_window]

                # all_jobs[:] = [job for job in all_jobs if job.submit_time > self.current_timestep + batch_window]

                cutoff = self.current_timestep + batch_window

                r = bisect_right(submit_times, cutoff, lo=cursor)

                if r > cursor:

                    jobs.extend(all_jobs[cursor:r])

                    cursor = r

            # 1. Prepare Timestep:

            completed_jobs, killed_jobs, newly_downed_nodes, need_reschedule = \

            simulation_done = self.complete_timestep(actively_considered_jobs=jobs,

                                                     all_jobs=all_jobs,

                                                     replay=replay,

                                                     autoshutdown=autoshutdown)

                                                     autoshutdown=autoshutdown,

                                                     cursor=cursor)

            if simulation_done:

                break

            yield tick_data

                if n not in self.available_nodes:

                    self.available_nodes.append(n)

                else:

                    raise KeyError(f"node was free but already in available nodes: {n.id}")

                    raise KeyError((f"Atempting to free node {n} after completion of job {job.id}. " +

                                     "Node is already free (in available nodes)!"))

            self.available_nodes = sorted(self.available_nodes)

    def update_system_utilization(self, current_time, running_jobs):

import pandas as pd

import sys

import os

import zmq

from ..policy import PolicyType, BackfillType

from raps.telemetry import Telemetry

from ..job import JobState

from raps.sim_config import args

from raps.system_config import get_system_config

# Run with this command:

# python main.py --system kestrel -f ../data/fastsim_jobs_output.parquet --scheduler fastsim --policy priority --start 2024-09-01T00:00 --end 2024-09-15T00:00

class Scheduler():

    """

    FastSim-backed scheduler (strict lockstep via ZeroMQ).

    Protocol (server side is FastSim --serve):

      - INIT                       -> { init_time }

      - GET { t }                  -> { t, running_ids }  (server acks t after reply)

      - END (on shutdown)          -> { ok: true }

    Semantics at engine second t:

      - R_t := authoritative running IDs from FastSim for t

      - started  = R_t - prev_R

        -> stamp start_time=t (once), assign nodes once, mark RUNNING

      - finished = prev_R - R_t

        -> stamp end_time=t (engine will finalize next tick in prepare_timestep)

      running list for this tick = R_t & finished  (so those finishing at t remain

      visible for one more scheduler call; engine completes them on next second).

    """

    def __init__(self, config, resource_manager, **kwargs):

        self.config = config

        self.policy = PolicyType(kwargs.get('policy'))

        self.bfpolicy = BackfillType(kwargs.get('backfill'))

        self.debug = bool(kwargs.get('debug', False))

        # ---- ZeroMQ client ----

        self.endpoint = kwargs.get('plugin_endpoint', 'ipc:///tmp/fastsim.sock')

        self._ctx = zmq.Context.instance()

        self._sock = self._ctx.socket(zmq.REQ)

        self._sock.setsockopt(zmq.LINGER, 0)

        self._sock.connect(self.endpoint)

        # INIT handshake: fetch FastSim's init_time (ISO string).

        self.init_time_iso = self._rpc('INIT').get('init_time')

        self.resource_manager = resource_manager

        # Job metadata: id -> Job

        self.jobids_to_jobs = {}

        self.allocated_jobs = set()   # job_ids we have assigned nodes for

        self.prev_running_ids = set() # R_{t-1}

        # Build the Job objects from RAPS Telemetry (needed so ExaDigiT subsystems have objects)

        args_dict = vars(args)

        config = get_system_config(args.system).get_legacy()

        args_dict['config'] = config

        td = Telemetry(**args_dict)

        print("...Now loading jobs to FastSim scheduler.")

        jobs, _, _ = td.load_data(args.replay)

        for job in jobs:

            self.jobids_to_jobs[job.id] = job

        if self.debug:

            print(f"[RAPS-FastSim] Connected to {self.endpoint}; init_time={self.init_time_iso}", file=sys.stderr)

    def _rpc(self, op, **payload):

        """Send a JSON request and return the JSON reply (dict)."""

        try:

            msg = {'op': op}

            msg.update(payload)

            self._sock.send_json(msg)

            rep = self._sock.recv_json()

        except Exception as e:

            raise RuntimeError(f"[RAPS-FastSim] RPC {op} failed: {e}") from e

        if isinstance(rep, dict) and 'error' in rep:

            raise RuntimeError(f"[RAPS-FastSim] RPC {op} error: {rep['error']}")

        return rep

    def _fastsim_running_ids(self, t: int):

        """Blocking call: get authoritative running job IDs for second t."""

        rep = self._rpc('GET', t=int(t))

        rids = rep.get('running_ids', [])

        return set(rids)

    def schedule(self, queue=None, running=None, current_time=None, accounts=None, sorted=False):

        """

        Called by Engine when RAPS detects an event.

        """

        running = running if running is not None else []

        t = int(current_time)

        # Get authoritative running set for second t (blocks until available)

        R_t = self._fastsim_running_ids(t)

        # Diff vs previous second

        started_ids  = R_t - self.prev_running_ids

        finished_ids = self.prev_running_ids - R_t  # these end at t; engine finalizes next tick

        # Handle starts: stamp start_time, assign nodes, mark RUNNING

        for jid in started_ids:

            job = self.jobids_to_jobs.get(jid)

            if job is None:

                if self.debug:

                    print(f"[RAPS-FastSim][WARN] Unknown job id from FastSim: {jid}", file=sys.stderr)

                continue

            # Assign nodes exactly once

            if jid not in self.allocated_jobs:

                self.resource_manager.assign_nodes_to_job(job, t, self.policy)

                self.allocated_jobs.add(jid)

            # FastSim is authoritative

            job.start_time = t

            # IMPORTANT: prevent premature completion by RM’s default behavior

            job.end_time = None # Prevents RAPS from removing job

            job.state = JobState.RUNNING

        # Handle finishes: stamp end_time=t (engine.prepare_timestep next tick completes)

        running.clear()

        for jid in finished_ids:

            job = self.jobids_to_jobs.get(jid)

            if job is not None:

                # overwrite any prior value; FastSim is the source of truth

                # job.end_time = t

                if job.start_time is not None:

                    observed = t - job.start_time

                    if (job.time_limit is None) or (job.time_limit < observed):

                        # This is necessary since RAPS is handling finishing jobs, but schedule is not always

                        # called at every tick, even though the job may have finished in FastSim during that tick.

                        # TODO: Deal with this, because it messes up the end time of some jobs.

                        # print(f"Extending {job.id} runtime {job.time_limit} to match observed {observed} at finish.")

                        job.time_limit = observed

                # print((f"Job {job.id} is finished, start time: {job.start_time}, wall time: {job.time_limit},"  

                #        f"end time: {job.end_time}, at time {t}. With nodes {job.scheduled_nodes}."))

                job.end_time = t

                job.time_limit = t - job.start_time

                running.append(job)

        # Running list reflects exactly FastSim’s R_t

        for jid in R_t:

            job = self.jobids_to_jobs.get(jid)

            if job is not None:

                # defensively ensure state isn’t stuck at COMPLETED

                if job.state != JobState.RUNNING:

                    job.state = JobState.RUNNING

                running.append(job)

        # Update prev

        self.prev_running_ids = R_t

    def end_sim(self):

        # Ask server to stop

        try:

            self._rpc('END')

        except Exception:

            pass

 No newline at end of file