File size: 15,178 Bytes
1a4aa87 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 | """
GPU-Aware Job Scheduler with Least-Load Assignment and Tenant Fairness
Provides intelligent job scheduling with:
- GPU affinity management
- Least-load worker selection
- Tenant fairness (preventing starvation)
- Atomic job claiming
- Fault tolerance
"""
import uuid
from datetime import datetime
from typing import Dict, List, Optional
from sqlalchemy import select, update, func
from backend.db.models import Worker, EvaluationRun
from backend.db.session import get_db_context
from backend.logging.logger import get_logger
from .job_schema import (
GPURequirement,
JobStatus,
EvaluationJob,
)
from .worker_registry import get_worker_registry, DEFAULT_HEARTBEAT_TIMEOUT
from .worker_schema import WorkerStatus
logger = get_logger("queue.scheduler", component="queue")
class JobScheduler:
"""
GPU-aware job scheduler with least-load assignment and tenant fairness.
Responsibilities:
- GPU affinity management
- Least-load worker selection
- Tenant fairness (weighted scheduling)
- Atomic job claiming
- Job assignment with capacity checking
"""
def __init__(self):
self._worker_registry = get_worker_registry()
async def get_tenant_active_job_count(self, tenant_id: uuid.UUID) -> int:
"""
Get the number of active jobs for a tenant.
Args:
tenant_id: The tenant ID
Returns:
Number of active jobs (pending or running)
"""
try:
async with get_db_context() as session:
query = select(func.count(EvaluationRun.id)).where(
EvaluationRun.tenant_id == tenant_id,
EvaluationRun.status.in_(["pending", "running"])
)
result = await session.execute(query)
return result.scalar() or 0
except Exception as e:
logger.error(
"Failed to get tenant active job count",
tenant_id=str(tenant_id),
error=str(e),
)
return 0
async def get_all_tenant_job_counts(self) -> Dict[uuid.UUID, int]:
"""
Get active job counts for all tenants.
Returns:
Dictionary mapping tenant_id to active job count
"""
try:
async with get_db_context() as session:
query = (
select(EvaluationRun.tenant_id, func.count(EvaluationRun.id))
.where(EvaluationRun.status.in_(["pending", "running"]))
.group_by(EvaluationRun.tenant_id)
)
result = await session.execute(query)
return {row[0]: row[1] for row in result.all()}
except Exception as e:
logger.error(
"Failed to get all tenant job counts",
error=str(e),
)
return {}
def calculate_tenant_priority(self, tenant_id: uuid.UUID, tenant_job_counts: Dict[uuid.UUID, int]) -> float:
"""
Calculate priority for a tenant based on job count.
Priority = 1 / (active_jobs_per_tenant + 1)
This gives higher priority to tenants with fewer active jobs,
preventing starvation.
Args:
tenant_id: The tenant ID
tenant_job_counts: Dictionary of tenant job counts
Returns:
Priority score (higher is better)
"""
job_count = tenant_job_counts.get(tenant_id, 0)
# Add 1 to avoid division by zero and give new tenants highest priority
return 1.0 / (job_count + 1)
async def get_pending_jobs_with_tenant_fairness(
self,
jobs: List[EvaluationJob],
) -> List[EvaluationJob]:
"""
Sort pending jobs by tenant fairness priority.
Jobs from tenants with fewer active jobs get higher priority.
Args:
jobs: List of pending jobs
Returns:
Sorted list of jobs
"""
if not jobs:
return jobs
# Get active job counts for all tenants
tenant_job_counts = await self.get_all_tenant_job_counts()
# Sort by tenant priority (highest priority first)
def get_priority(job: EvaluationJob) -> float:
if hasattr(job, 'tenant_id') and job.tenant_id:
return self.calculate_tenant_priority(job.tenant_id, tenant_job_counts)
return 0.0 # Jobs without tenant get lowest priority
return sorted(jobs, key=get_priority, reverse=True)
async def assign_job_to_worker(
self,
job: EvaluationJob,
) -> Optional[str]:
"""
Assign a job to the best available worker using least-load strategy.
The algorithm:
1. Filter workers by GPU requirement
2. Filter workers by status (ACTIVE or DEGRADED)
3. Filter workers with capacity (active_jobs < max_concurrent_jobs)
4. Sort by load factor (active_jobs / max_concurrent_jobs)
5. Select the worker with lowest load factor
Args:
job: The job to assign
Returns:
Worker ID if assigned, None if no suitable worker found
"""
try:
# Determine GPU requirement from job
gpu_required = self._get_gpu_requirement(job)
# Get available workers
available_workers = await self._worker_registry.get_available_workers(
gpu_required=gpu_required
)
if not available_workers:
logger.warning(
"No available workers for job",
job_id=str(job.job_id),
gpu_required=gpu_required,
)
return None
# Select worker with least load
selected_worker = None
min_load = float('inf')
for worker in available_workers:
# Calculate load factor
if worker.max_concurrent_jobs > 0:
load_factor = worker.active_jobs / worker.max_concurrent_jobs
else:
load_factor = float('inf')
# Check GPU capacity if GPU required
if gpu_required > 0:
# Check if worker has enough free GPU memory
free_gpu_memory = worker.gpu_memory_total - worker.gpu_memory_used
if free_gpu_memory < 4000: # Require at least 4GB free per job
continue
if load_factor < min_load:
min_load = load_factor
selected_worker = worker
if selected_worker is None:
logger.warning(
"No worker with sufficient capacity",
job_id=str(job.job_id),
)
return None
# Atomically claim the job
worker_id = await self._claim_job_for_worker(
job_id=job.job_id,
worker_id=selected_worker.worker_id,
)
if worker_id:
logger.info(
"Job assigned to worker",
job_id=str(job.job_id),
worker_id=worker_id,
load_factor=min_load,
)
return worker_id
except Exception as e:
logger.error(
"Failed to assign job to worker",
job_id=str(job.job_id),
error=str(e),
)
return None
async def _claim_job_for_worker(
self,
job_id: uuid.UUID,
worker_id: str,
) -> Optional[str]:
"""
Atomically claim a job for a worker.
Uses atomic UPDATE to prevent duplicate job execution.
Args:
job_id: Job ID
worker_id: Worker ID
Returns:
Worker ID if claimed successfully, None if already claimed
"""
try:
from backend.queue.producer import _job_queue
# Find the job in the queue
job = None
for j in _job_queue:
if j.job_id == job_id:
job = j
break
if job is None:
logger.warning(
"Job not found in queue",
job_id=str(job_id),
)
return None
# Check if job is still queued (not already claimed)
if job.status != JobStatus.QUEUED:
logger.warning(
"Job not in QUEUED status",
job_id=str(job_id),
status=job.status,
)
return None
# Atomically update job status and worker
job.status = JobStatus.RUNNING
job.worker_id = worker_id
job.started_at = datetime.utcnow()
# Update worker active jobs count
async with get_db_context() as session:
stmt = (
update(Worker)
.where(Worker.worker_id == worker_id)
.values(active_jobs=Worker.active_jobs + 1)
)
await session.execute(stmt)
await session.commit()
logger.debug(
"Job claimed atomically",
job_id=str(job_id),
worker_id=worker_id,
)
return worker_id
except Exception as e:
logger.error(
"Failed to claim job",
job_id=str(job_id),
worker_id=worker_id,
error=str(e),
)
return None
def _get_gpu_requirement(self, job: EvaluationJob) -> int:
"""
Determine GPU requirement from job.
Args:
job: The job
Returns:
Number of GPUs required (0 for CPU-only)
"""
# Check job metadata for GPU requirement
if job.metadata:
gpu_req = job.metadata.get("gpu_requirement")
if gpu_req is not None:
return int(gpu_req)
# Infer from job type
job_type = job.metadata.get("job_type")
if job_type == "benchmark":
return 1 # Benchmark jobs typically need GPU
elif job_type == "single_eval":
return 0 # Single eval can run on CPU
# Default to 1 GPU for benchmark jobs
if hasattr(job, 'job_type') and job.job_type == "benchmark":
return 1
return 0
async def release_job_from_worker(
self,
job_id: uuid.UUID,
worker_id: str,
) -> bool:
"""
Release a job from a worker (job completed or failed).
Args:
job_id: Job ID
worker_id: Worker ID
Returns:
True if released successfully
"""
try:
# Update worker active jobs count
async with get_db_context() as session:
stmt = (
update(Worker)
.where(Worker.worker_id == worker_id)
.where(Worker.active_jobs > 0)
.values(active_jobs=Worker.active_jobs - 1)
)
await session.execute(stmt)
await session.commit()
logger.debug(
"Job released from worker",
job_id=str(job_id),
worker_id=worker_id,
)
return True
except Exception as e:
logger.error(
"Failed to release job from worker",
job_id=str(job_id),
worker_id=worker_id,
error=str(e),
)
return False
async def get_worker_for_job(
self,
gpu_required: int = 0,
) -> Optional[Worker]:
"""
Get the best worker for a job with given GPU requirements.
Args:
gpu_required: Number of GPUs required
Returns:
Best worker or None
"""
available_workers = await self._worker_registry.get_available_workers(
gpu_required=gpu_required
)
if not available_workers:
return None
return available_workers[0] # Already sorted by load factor
async def check_gpu_capacity(
self,
worker_id: str,
gpu_required: int,
) -> bool:
"""
Check if a worker has sufficient GPU capacity for a job.
Args:
worker_id: Worker ID
gpu_required: GPUs required
Returns:
True if worker has sufficient capacity
"""
try:
async with get_db_context() as session:
stmt = select(Worker).where(Worker.worker_id == worker_id)
result = await session.execute(stmt)
worker = result.scalar_one_or_none()
if worker is None:
return False
# Check GPU count
if worker.gpu_count < gpu_required:
return False
# Check GPU memory
free_memory = worker.gpu_memory_total - worker.gpu_memory_used
required_memory = gpu_required * 4000 # 4GB per GPU minimum
return free_memory >= required_memory
except Exception as e:
logger.error(
"Failed to check GPU capacity",
worker_id=worker_id,
error=str(e),
)
return False
# Global instance
_scheduler: Optional[JobScheduler] = None
def get_job_scheduler() -> JobScheduler:
"""Get the global job scheduler instance."""
global _scheduler
if _scheduler is None:
_scheduler = JobScheduler()
return _scheduler
__all__ = [
"JobScheduler",
"get_job_scheduler",
]
|