Performetrica HPC – Getting Started & System Overview

Documentation Status

Accessing Performetrica HPC

From inside Performetrica / local network

ssh <username>@login.superbilgisayar.tr

From outside (internet)

ssh <username>@login.superbilgisayar.tr

If VPN or firewall restrictions apply, follow Performetrica access policy.

Cluster Overview

Performetrica HPC is a CPU-based high performance computing cluster designed for:

  • Scientific computing
  • Parallel workloads (MPI / OpenMP)
  • Simulation, modeling, and data processing

Inspecting Cluster Resources

To list nodes and partitions:

sinfo --long --Node "%#N %.6D %#P %6t"

To check detailed node info:

scontrol show node <nodename>

Understanding CPU Capabilities

Performance depends heavily on CPU features.

Example CPU capabilities (Intel Xeon class CPUs):

  • AVX / AVX2 / AVX-512
  • FMA (Fused Multiply-Add)
  • SIMD vectorization
  • NUMA architecture

You can inspect CPU flags:

lscpu

or:

cat /proc/cpuinfo

Why this matters

  • AVX/AVX512 → faster vector math
  • NUMA → memory locality is critical
  • Cache hierarchy → affects scaling

Performance Optimization Guidelines

1. Match workload to architecture

  • Use OpenMP for shared memory scaling
  • Use MPI for distributed scaling

2. CPU Binding (Critical)

Avoid CPU migration:

#SBATCH --cpu-bind=cores

3. NUMA Awareness

  • Keep threads on same socket if possible
  • Avoid cross-socket memory traffic

4. Thread Configuration

export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK

5. Memory Optimization

  • Request realistic memory
  • Avoid over-allocation (increases queue time)
  • Monitor usage with:
sacct -j <jobid> --format=MaxRSS

6. I/O Considerations

  • Use local /tmp when possible
  • Avoid heavy I/O on shared storage
  • Batch writes instead of frequent small writes

Typical Workload Types

Workload Type Recommended Slurm Settings
Serial ntasks=1, cpus-per-task=1
OpenMP ntasks=1, cpus-per-task=N
MPI ntasks=N
Hybrid ntasks + cpus-per-task

Example CPU Job

#!/bin/bash
#SBATCH --job-name=test
#SBATCH --partition=defq
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=16
#SBATCH --mem=32G
#SBATCH --time=02:00:00

export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK

./my_application

Key Takeaways

  • Use correct Slurm parameters
  • Respect NUMA and cache locality
  • Optimize total runtime (queue + execution)

References

  • https://slurm.schedmd.com/
  • https://hpc-wiki.info/