CraneSched vs. Slurm Feature Comparison
CraneSched is China's first open-source, domestically-developed compute scheduling system supporting both HPC and AI workloads. It comprehensively benchmarks against Slurm — the leading international scheduler — and surpasses it in performance, container support, and domestic hardware compatibility.
CraneSched significantly outperforms Slurm and OpenPBS in scheduling throughput. Measured results:
Average Jobs Scheduled Per Minute
Scheduler
Avg. Jobs/Min
Relative to CraneSched
CraneSched 105,538 1x
OpenPBS
11,136
9.4x slower
Slurm
4,259
24.7x slower
Peak Jobs Scheduled Per Minute
Scheduler
Peak Jobs/Min
Relative to CraneSched
CraneSched 122,427 1x
OpenPBS
20,541
6x slower
Slurm
4,551
26.9x slower
Metric
CraneSched
Scheduling throughput
5–20x faster than Slurm
Cluster scale
Supports 100,000+ nodes
Job throughput
Real-time scheduling of 10,000+/sec ; hourly throughput exceeds 38 million
Concurrency
2,000,000+ concurrent jobs
Response latency
Millisecond-level low latency
Scheduling Feature Comparison
Feature
CraneSched
Slurm
Description
Basic Scheduling
Backfill Scheduling
Run short jobs in idle time windows to improve utilization
Fair-Share Scheduling
Fair scheduling policy based on historical usage
Priority Scheduling
Multi-factor priority calculation
FIFO Scheduling
Basic first-in, first-out scheduling
Resource Management
Preemption
High-priority jobs preempt resources from lower-priority ones
Reservation
Reserve resource time windows for specific users or jobs
TRES Fine-Grained Tracking
Trackable resource types (CPU, memory, GPU, etc.)
QOS Management
Differentiated service level control
Resource Escape Protection
Prevent jobs from exceeding allocated resources
Job Management
Job Dependencies
Control dependency relationships between jobs
Job Arrays
Batch submission of parameterized jobs
Job Steps
Multi-step management within a job
Interactive Jobs
Real-time interactive computing
Energy Saving & Efficiency
Power Saving Scheduling
Automatically shut down idle nodes under low load
AI Job Runtime Prediction (ORA)
LLM-based job runtime prediction; 41% accuracy improvement
Smart Fair-Share (TSMF)
In-house two-stage multi-factor algorithm; utilization improved to 97.3%
Automated Power Saving (EcoSched)
Automated power control; 78.64% energy reduction under low load
Account & Permissions
Hierarchical Account Management
Tree-structured user/account management
RBAC Access Control
Role-based access control
High Availability
Automatic Fault Recovery
Automatic recovery after control node failure
Distributed Fault Tolerance
No single point of failure
Container Support Comparison
CraneSched and Slurm take different technical approaches to container support:
Dimension
CraneSched
Slurm
Technical approach Based on K8s-underlying CRI RPC interface (the de facto cloud-native standard)
OCI model compatibility via CLI
Container runtime containerd / CRI-O (via CRI interface)
Singularity / Enroot (via CLI)
Image management Auto-pull; no manual handling needed
User must download and convert image formats
Dedicated CLI ccon command, Docker CLI-inspired design, easy to learnNative Slurm commands (sbatch/srun), different from Docker usage
Network isolation CNI-based multi-tenant network isolation (Calico Underlay)
Minimal support
Filesystem isolation Full user/network/mount namespace isolation
Limited
Fake Root User Namespace-based, root experience inside container
Relies on Singularity's fakeroot
RDMA support Supports SR-IOV shared RNIC and direct passthrough
Limited
Operations tools Mature tools: crictl/nerdctl/ctr, etc.
Relies on community tools
Unique Advantages of CraneSched Containers
Ease of use : No manual image pulling; dedicated CLI (ccon) designed for Docker users with no Slurm experienceComplete network isolation : CNI support allows admins to implement various container networking strategies including multi-tenant isolationRDMA network support : Supports mid-to-large-scale RoCE networks (SR-IOV) and large-scale AI training clusters (Spine-Leaf architecture)Pod/Job concept mapping : Maps K8s Pod/Container concepts to Job/Step, enabling imperative orchestration
Command Compatibility
CraneSched provides an in-house Slurm & LSF Wrapper with full compatibility for Slurm and LSF command-line syntax:
Slurm Command
CraneSched Native
Function
sbatchcbatchSubmit batch jobs
squeuecqueueView job queue
sruncrunRun interactive jobs
salloccallocAllocate interactive resources
sinfocinfoView cluster information
sacctcacctView job history
sacctmgrcacctmgrAccount management
scancelccancelCancel jobs
scontrolccontrolSystem control
Zero migration cost : Via the Slurm Wrapper, users can switch from Slurm to CraneSched transparently without modifying any scripts or workflows. Peking University's Weiming Teaching Cluster No.2 has successfully completed a transparent migration from Slurm to CraneSched, supporting hundreds of user software packages.
Heterogeneous Hardware Support
CraneSched fully supports mainstream domestic and international hardware platforms:
Architecture Support
Architecture
Support
X86
ARM
RISC-V
CPU Compatibility
Category
Supported Brands
International
Intel, AMD
Domestic
Phytium, Hygon, Huawei Kunpeng
Accelerator Compatibility
Category
Supported Brands
International
Nvidia GPU, AMD GPU
Domestic
Huawei Ascend, Hygon DCU, Cambricon MLU, Iluvatar CoreX, Kunlunxin, Metax, Moore Threads
Operating System Compatibility
Category
Supported Systems
International
CentOS, Ubuntu, Rocky Linux
Domestic
OpenEuler, KylinOS
CraneSched has received product compatibility certifications from multiple vendors including Inspur, Phytium, Hygon, and Kunlunxin.
Summary
Dimension
CraneSched Advantage
Performance 5–20x faster than Slurm
Features Full Slurm feature coverage plus AI prediction, intelligent power saving, and more
Containers Native CRI/CNI support, multi-tenant network isolation, RDMA support
Compatibility Fully compatible with Slurm/LSF commands, zero migration cost
Domestic hardware Full support for domestic CPUs, GPUs/NPUs, and operating systems
Convergence HPC + AI integration; full Storage·Compute·Usage convergence