Describe characteristics of mainframe computers and supercomputers

Hardware & Software – Mainframe Computers and Supercomputers (Cambridge AS & A‑Level IT 9626)

1. Mainframe Computers – Core Characteristics

Large, high‑capacity systems that run continuously for mission‑critical business applications such as banking, government, and large‑scale enterprise processing.

• Longevity & Upgrade Path
  • Designed for a service life of 10‑20 years.
  • Vendors supply regular hardware refreshes (CPU, memory, I/O) that can be installed online, without shutting the system down.
• RAS – Reliability, Availability, Serviceability
  • Reliability: Mean Time Between Failures (MTBF) measured in years; redundant power supplies, fans and I/O channels.
  • Availability: Target uptime of 99.999 % (five‑nines) via hot‑swap modules and automatic fail‑over.
  • Serviceability: On‑line diagnostics, predictive failure analysis, and easy replacement of faulty parts.
• Fault‑tolerance & Redundancy
  • Dual‑system (active‑active) configurations.
  • Error‑correcting code (ECC) memory.
  • Built‑in checkpoint/restart for batch jobs.
• Security
  • Hardware cryptographic coprocessors and secure boot.
  • Role‑based access control, pervasive encryption, and immutable audit trails.
• Performance Metrics
  • Typical throughput: 10‑100 MIPS per processor; modern cores exceed 200 MIPS.
  • Floating‑point capability: up to several hundred GFLOPS (e.g., IBM z15 ≈ 200 GFLOPS).
  • Benchmarks: TPC‑C (transaction processing), SPECjbb (Java business‑logic), and LINPACK for FLOPS.
• I/O Volume
  • Millions of I/O operations per second via 64‑bit FICON, ESCON, or NVMe‑over‑Fabrics channels.
  • Massive storage arrays – petabytes of DAS/DSS with sub‑microsecond latency.
• Processor Count & Architecture
  • From 2 up to 64+ high‑performance CPUs (e.g., IBM z15 up to 190 cores, each with multiple hardware threads).
  • RISC‑based designs with deep pipelines and large caches.
• Memory Capacity
  • Up to 190 TB of shared main memory (IBM z15) with optional tiered cache.
• Heat Management
  • Advanced liquid‑cooling and hot‑air containment.
  • Thermal sensors drive automatic fan‑speed adjustment and throttling.
• Operating Systems
  • IBM z/OS, Linux on IBM Z, z/VM (virtualisation), z/VSE.
  • Unisys OS 2200, ClearPath OS.
• Virtualisation
  • Logical Partitions (LPARs) and containers allow many isolated OS instances on a single chassis.
• Multi‑user Support
  • Thousands of concurrent terminals, batch jobs, and online transaction processing (OLTP) workloads.

Disadvantages (syllabus‑required)

• Very high acquisition and ongoing maintenance cost.
• Complex configuration; requires specialist staff for installation and management.
• Vendor lock‑in – proprietary hardware, firmware and operating systems.
• Physical size and power consumption (hundreds of kW to a few MW).

2. Supercomputers – Core Characteristics

Specialised high‑performance machines built to solve the world’s most demanding scientific, engineering and data‑intensive problems.

• Extreme Computational Speed
  • Measured in FLOPS; current top systems achieve 0.4‑1 exaflop (10¹⁸ FLOPS) peak performance.
• Massive Parallelism
  • Thousands to millions of CPU cores, often combined with GPU, FPGA or AI‑accelerator cards.
• Specialised Architecture
  • High‑speed interconnects (InfiniBand HDR, custom torus, Omni‑Path) with sub‑microsecond latency.
  • Node‑level memory bandwidth > 1 TB/s in modern designs.
• Energy Consumption & Cooling
  • Power budgets of 5‑15 MW; require liquid‑cooling, chilled‑water loops or immersion cooling.
• Custom Software Stack
  • Optimised Linux kernels (e.g., Cray Linux Environment, SUSE Linux Enterprise HPC).
  • Parallel programming libraries: MPI, OpenMP, CUDA, OpenACC.
  • Specialised compilers (Intel OneAPI, GNU, PGI) that auto‑vectorise for SIMD units.
• Application‑Specific Optimisation
  • Code is often rewritten to exploit node topology, memory hierarchy and accelerator features.
• Reliability Strategies
  • Checkpoint/restart and resilient MPI are the primary techniques for tolerating node failures.
  • Occasional hardware faults are accepted; jobs are automatically rescheduled from the last checkpoint.
• Typical Users
  • Researchers, scientists, engineers, climate modelers, data‑intensive analysts.
• Refresh Cycle
  • Major hardware upgrades are normally required every 5‑8 years to stay competitive.

Disadvantages (syllabus‑required)

• Enormous capital and operating costs (hardware, electricity, cooling staff).
• Optimised for batch‑oriented, highly parallel workloads; not suitable for general‑purpose business applications.
• Complex software environment; requires specialised programming expertise.

3. Advantages & Disadvantages – Comparative Table

4. Direct Comparison – Quick Reference

5. Example Systems (2025)

1. IBM z15 – Mainframe with up to 190 TB memory, 190 cores, integrated cryptographic coprocessors; runs z/OS, Linux on Z, and z/VM.
2. IBM Power10 E‑Series – Power‑based mainframe, up to 256 cores, 4 TB memory per drawer, AI‑enhanced transaction processing.
3. Fugaku (RIKEN, Japan) – ARM‑based supercomputer; peak 442 PFLOPS (Rmax 256 PFLOPS), 7.6 million cores, 30 MW power envelope.
4. Summit (Oak Ridge, USA) – IBM Power9 + NVIDIA V100 GPUs; peak 200 PFLOPS, 4.5 MW, 9,216 nodes.
5. Frontier (Oak Ridge, USA) – First exascale system; AMD EPYC CPUs + Radeon Instinct GPUs; 1.1 exaflop peak, 21 MW.

6. Summary

Both mainframe computers and supercomputers occupy the top tier of computing technology, but they serve fundamentally different purposes. Mainframes prioritise continuous availability, massive I/O throughput, robust security and a long service life for enterprise workloads. Supercomputers maximise raw parallel processing power, inter‑node bandwidth and specialised software to tackle scientific challenges that require exascale performance. Understanding the distinct characteristics, advantages and drawbacks of each system enables students to choose the appropriate architecture for a given problem domain and to appreciate the engineering trade‑offs that underpin modern high‑performance computing.