iowarp-core 1.5.8

IOWarp Context Management Platform

CLIO Core

A Comprehensive Platform for Context Management in Scientific Computing

Overview · Installation · Components · Quickstart · Documentation · Contributing

---

Overview

CLIO Core is a unified framework that integrates multiple high-performance components for context management, data transfer, and scientific computing. Built with a modular architecture, CLIO Core enables developers to create efficient data processing pipelines for HPC, storage systems, and near-data computing applications.

CLIO Core provides:

• High-Performance Context Management: Efficient handling of computational contexts and data transformations
• Heterogeneous-Aware I/O: Multi-tiered, dynamic buffering for accelerated data access
• Modular Runtime System: Extensible architecture with dynamically loadable processing modules
• Advanced Data Structures: Shared memory compatible containers with GPU support (CUDA, ROCm)
• Distributed Computing: Seamless scaling from single node to cluster deployments

Architecture

CLIO Core follows a layered architecture integrating five core components:

┌──────────────────────────────────────────────────────────────┐
│                      Applications                            │
│          (Scientific Workflows, HPC, Storage Systems)        │
└──────────────────────────────────────────────────────────────┘
                              │
        ┌─────────────────────┼─────────────────────┐
        │                     │                     │
┌───────────────┐   ┌──────────────────┐   ┌────────────────┐
│   Context     │   │    Context       │   │   Context      │
│  Exploration  │   │  Assimilation    │   │   Transfer     │
│    Engine     │   │     Engine       │   │    Engine      │
└───────────────┘   └──────────────────┘   └────────────────┘
        │                     │                     │
        └─────────────────────┼─────────────────────┘
                              │
                    ┌─────────────────┐
                    │  Chimaera       │
                    │  Runtime        │
                    │  (Module System)│
                    └─────────────────┘
                              │
                ┌─────────────────────────┐
                │  Context Transport      │
                │  Primitives             │
                │  (Shared Memory & IPC)  │
                └─────────────────────────┘

Installation

Pip (recommended)

The pip wheel is the easiest way to get CLIO Core. It ships a portable, self-contained build with all dependencies statically linked. No system installs are required beyond glibc and Python 3.10+.

pip install iowarp-core

The wheel includes the Clio runtime, the chimaera CLI, the CTE, CAE, and CEE engines, and the clio_cee Python bindings. A default configuration is seeded at ~/.clio/clio.yaml (legacy: ~/.chimaera/chimaera.yaml) on first import.

Newer extensions and advanced/accelerated features are not in the portable wheel — switch to a source build below if you need any of:

• NVIDIA GPU (CUDA) or AMD GPU (ROCm) acceleration
• MPI for distributed multi-node deployment
• HDF5 / ADIOS2 adapters
• FUSE adapter
• Compression backends (LibPressio, Blosc, etc.)
• Custom ChiMods built against the C++ headers
• Sanitizer or debug builds

Source build (Conda)

git clone --recurse-submodules https://github.com/iowarp/clio-core.git
cd clio-core
bash install.sh release

release corresponds to a variant stored under installers/conda/variants/. Other variants (cuda, rocm, mpi, full, release-fuse, debug, ...) enable the corresponding features. Feel free to add a new variant for your specific machine there.

If you already cloned without submodules, initialize them with:

git submodule update --init --recursive

Other source-install methods (Docker, Spack) are documented in docs/getting-started/installation.

Components

CLIO Core consists of five integrated components, each with its own specialized functionality:

1. Context Transport Primitives

Location: context-transport-primitives/

High-performance shared memory library containing data structures and synchronization primitives compatible with shared memory, CUDA, and ROCm.

Key Features:

• Shared memory compatible data structures (vector, list, unordered_map, queues)
• GPU-aware allocators (CUDA, ROCm)
• Thread synchronization primitives
• Networking layer with ZMQ transport
• Compression and encryption utilities

Read more →

2. Chimaera Runtime

Location: context-runtime/

High-performance modular runtime for scientific computing and storage systems with coroutine-based task execution.

Key Features:

• Ultra-high performance task execution (< 10μs latency)
• Modular Module system for dynamic extensibility
• Coroutine-aware synchronization (CoMutex, CoRwLock)
• Distributed architecture with shared memory IPC
• Built-in storage backends (RAM, file-based, custom block devices)

Read more →

3. Context Transfer Engine

Location: context-transfer-engine/

Heterogeneous-aware, multi-tiered, dynamic I/O buffering system designed to accelerate I/O for HPC and data-intensive workloads.

Key Features:

• Programmable buffering across memory/storage tiers
• Multiple I/O pathway adapters
• Integration with HPC runtimes and workflows
• Improved throughput, latency, and predictability

Read more →

4. Context Assimilation Engine

Location: context-assimilation-engine/

High-performance data ingestion and processing engine for heterogeneous storage systems and scientific workflows.

Key Features:

• OMNI format for YAML-based job orchestration
• MPI-based parallel data processing
• Binary format handlers (Parquet, CSV, custom formats)
• Repository and storage backend abstraction
• Integrity verification with hash validation

Read more →

5. Context Exploration Engine

Location: context-exploration-engine/

Interactive tools and interfaces for exploring scientific data contents and metadata.

Key Features:

• Model Context Protocol (MCP) for HDF5 data
• HDF Compass viewer (wxPython-4 based)
• Interactive data exploration interfaces
• Metadata browsing capabilities

Read more →

Quickstart

Starting the Runtime

Installation seeds a default configuration at ~/.clio/clio.yaml (legacy: ~/.chimaera/chimaera.yaml), so the runtime works out of the box:

# Foreground
clio_run runtime start

# Background
clio_run runtime start &

To override the configuration, point CLIO_X at your YAML file:

export CLIO_X=/path/to/my_config.yaml
clio_run runtime start

Context Exploration Engine Python Example

Here we show an example of how to use the context exploration engine to bundle and retrieve data.

import clio_cee as cee

# Create ContextInterface (handles runtime initialization internally)
ctx_interface = cee.ContextInterface()

# Assimilate a file into IOWarp storage
ctx = cee.AssimilationCtx(
    src="file::/path/to/data.bin",      # Source: local file
    dst="iowarp::my_dataset",            # Destination: IOWarp tag
    format="binary"                      # Format: binary, hdf5, etc.
)
result = ctx_interface.context_bundle([ctx])
print(f"Assimilation result: {result}")

# Query for blobs matching a pattern
blobs = ctx_interface.context_query(
    "my_dataset",    # Tag name
    ".*",            # Blob name regex (match all)
    0                # Flags
)
print(f"Found blobs: {blobs}")

# Retrieve blob data
packed_data = ctx_interface.context_retrieve(
    "my_dataset",    # Tag name
    ".*",            # Blob name regex
    0                # Flags
)
print(f"Retrieved {len(packed_data)} bytes")

# Cleanup when done
ctx_interface.context_destroy(["my_dataset"])

Context Transfer Engine C++ Example

Here is an example of the context transfer engine's C++ API.

#include <clio_cte/core/core_client.h>
#include <clio_runtime/clio_runtime.h>

int main() {
  // 1. Initialize Clio runtime
  bool success = chi::CHIMAERA_INIT(chi::ChimaeraMode::kClient, true);
  if (!success) return 1;

  // 2. Initialize CTE subsystem
  clio_cte::core::CLIO_CTE_CLIENT_INIT();

  // 3. Create CTE client
  clio_cte::core::Client cte_client;
  clio_cte::core::CreateParams params;
  cte_client.Create(chi::PoolQuery::Dynamic(),
                    clio_cte::core::kCtePoolName,
                    clio_cte::core::kCtePoolId, params);

  // 4. Register a storage target (100MB file-based)
  cte_client.RegisterTarget("/tmp/cte_storage",
                            chimaera::bdev::BdevType::kFile,
                            100 * 1024 * 1024);

  // 5. Create a tag (container for blobs)
  clio_cte::core::TagId tag_id = cte_client.GetOrCreateTag(
      "my_tag", clio_cte::core::TagId::GetNull());

  // 6. Store blob data
  std::vector<char> data(4096, 'A');
  ctp::ipc::FullPtr<char> shared_data = CHI_IPC->AllocateBuffer(data.size());
  memcpy(shared_data.ptr_, data.data(), data.size());

  cte_client.PutBlob(tag_id, "my_blob",
                     0,                    // offset
                     data.size(),          // size
                     shared_data.shm_,     // shared memory pointer
                     0.8f,                 // importance score
                     0);                   // flags
  CHI_IPC->FreeBuffer(shared_data);

  // 7. Retrieve blob data
  ctp::ipc::FullPtr<char> read_buf = CHI_IPC->AllocateBuffer(data.size());
  cte_client.GetBlob(tag_id, "my_blob",
                     0,                    // offset
                     data.size(),          // size
                     0,                    // flags
                     read_buf.shm_);
  // read_buf.ptr_ now contains the retrieved data
  CHI_IPC->FreeBuffer(read_buf);

  // 8. Cleanup
  cte_client.DelTag(tag_id);
  return 0;
}

Build and Link:

# Unified package includes everything - ClioCtp, Chimaera, and all ChiMods
find_package(iowarp-core REQUIRED)

target_link_libraries(my_app
  clio_cte::core_client    # CTE client (for the example above)
  chimaera::admin_client  # Admin Module (always available)
  chimaera::bdev_client   # Block device Module (always available)
)

What find_package(iowarp-core) provides:

Core Components:

• All ctp::* modular targets (cxx, configure, serialize, interceptor, lightbeam, thread_all, mpi, compress, encrypt)
• chimaera::cxx (core runtime library)
• Module build utilities

Core ChiMods (Always Available):

• chimaera::admin_client, chimaera::admin_runtime
• chimaera::bdev_client, chimaera::bdev_runtime

Optional ChiMods (if enabled at build time):

• clio_cte::core_client, clio_cte::core_runtime (Context Transfer Engine)
• clio_cae::core_client, clio_cae::core_runtime (Context Assimilation Engine)

Testing

CLIO Core includes comprehensive test suites for each component:

# Run all unit tests
cd build
ctest -VV

# Run specific component tests
ctest -R context_transport  # Transport primitives tests
ctest -R chimaera           # Runtime tests
ctest -R cte                # Context transfer engine tests
ctest -R omni               # Context assimilation engine tests

Benchmarking

CLIO Core includes performance benchmarks for measuring runtime and I/O throughput.

Runtime Throughput Benchmark (clio_run_thrpt_benchmark)

Measures task throughput and latency for the Clio runtime.

clio_run_thrpt_benchmark [options]

Parameters:

Parameter	Default	Description
--test-case <case>	bdev_io	Test case to run
--threads <N>	4	Number of client worker threads
--duration <seconds>	10.0	Duration to run benchmark
--max-file-size <size>	1g	Maximum file size (supports k, m, g suffixes)
--io-size <size>	4k	I/O size per operation
--lane-policy <P>	(from config)	Lane policy: map_by_pid_tid, round_robin, random
--output-dir <dir>	/tmp/clio_benchmark	Output directory for files
--verbose, -v	false	Enable verbose output

Test Cases:

• bdev_io - Full I/O throughput (Allocate → Write → Free)
• bdev_allocation - Allocation-only throughput
• bdev_task_alloc - Task allocation/deletion overhead
• latency - Round-trip task latency

Examples:

# Full I/O benchmark with 8 threads for 30 seconds
clio_run_thrpt_benchmark --test-case bdev_io --threads 8 --duration 30

# Latency benchmark with verbose output
clio_run_thrpt_benchmark --test-case latency --threads 4 --verbose

# Large I/O with 1MB blocks
clio_run_thrpt_benchmark --test-case bdev_io --io-size 1m --threads 16

CTE Benchmark (clio_cte_bench)

Measures Context Transfer Engine Put/Get performance.

clio_cte_bench [options]

Parameters:

Parameter	Default	Description
--op <Put|Get|PutGet>	Put	Operation to benchmark (alias: --test-case)
--threads <N>	1	Number of worker threads
--depth <N>	1	Async requests in flight per thread
--io-size <size>	1m	Bytes per op (supports k, m, g suffixes)
--io-count <N>	1000	Ops per thread (ignored when --time-limit is set)
--max-total-blobs <N>	0 (unbounded)	Total distinct keys across all threads (split evenly)
--time-limit <seconds>	0 (off)	Run for N seconds instead of --io-count ops
--help, -h		Show usage and exit

A legacy positional form is also accepted:

clio_cte_bench <test_case> <threads> <depth> <io_size> <io_count>

Examples:

# Put benchmark: 4 threads, 8 async depth, 1MB I/O, 200 ops/thread
clio_cte_bench --op Put --threads 4 --depth 8 --io-size 1m --io-count 200

# Get benchmark with a 30-second time limit and 4KB I/O
clio_cte_bench --op Get --threads 2 --depth 4 --io-size 4k --time-limit 30

# Combined Put/Get over a bounded keyspace of 1024 total blobs
clio_cte_bench --op PutGet --threads 8 --depth 16 --io-size 16m \
               --io-count 50 --max-total-blobs 1024

Output Metrics:

• Total execution time (ms)
• Per-thread bandwidth: min, max, avg (MB/s)
• Aggregate bandwidth across all threads

Documentation

Comprehensive documentation is available for each component:

• AGENTS.md: Unified development guide and coding standards
• Context Transport Primitives: Shared memory data structures
• Chimaera Runtime: Modular runtime system and Module development
  • MODULE_DEVELOPMENT_GUIDE.md: Complete Module development guide
• Context Transfer Engine: I/O buffering and acceleration
  • CTE API Documentation: Complete API reference
• Context Assimilation Engine: Data ingestion and processing
• Context Exploration Engine: Interactive data exploration

Use Cases

Scientific Computing:

• High-performance data processing pipelines
• Near-data computing for large datasets
• Custom storage engine development
• Computational workflows with context management

Storage Systems:

• Distributed file system backends
• Object storage implementations
• Multi-tiered cache and storage solutions
• High-throughput I/O buffering

HPC and Data-Intensive Workloads:

• Accelerated I/O for scientific applications
• Data ingestion and transformation pipelines
• Heterogeneous computing with GPU support
• Real-time streaming analytics

Performance Characteristics

CLIO Core is designed for high-performance computing scenarios:

• Task Latency: < 10 microseconds for local task execution (Chimaera Runtime)
• Memory Bandwidth: Up to 50 GB/s with RAM-based storage backends
• Scalability: Single node to multi-node cluster deployments
• Concurrency: Thousands of concurrent coroutine-based tasks
• I/O Performance: Native async I/O with multi-tiered buffering

Contributing

We welcome contributions to the CLIO Core project!

Development Workflow

1. Fork the repository
2. Create a feature branch: git checkout -b feature/amazing-feature
3. Follow the coding standards in AGENTS.md
4. Test your changes: ctest --test-dir build
5. Submit a pull request

Coding Standards

• Follow Google C++ Style Guide
• Use semantic naming for IDs and priorities
• Always create docstrings for new functions (Doxygen compatible)
• Add comprehensive unit tests for new functionality
• Never use mock/stub code unless explicitly required - implement real, working code

See AGENTS.md for complete coding standards and workflow guidelines.

License

CLIO Core is licensed under the BSD 3-Clause License. See LICENSE file for complete license text.

Copyright (c) 2024, Gnosis Research Center, Illinois Institute of Technology

---

Acknowledgements

CLIO Core is developed at the GRC lab at Illinois Institute of Technology as part of the IOWarp project. This work is supported by the National Science Foundation (NSF) and aims to advance next-generation scientific computing infrastructure.

For more information:

• IOWarp Project: https://grc.iit.edu/research/projects/iowarp
• IOWarp Organization: https://github.com/iowarp
• Documentation Hub: https://grc.iit.edu/docs/category/iowarp

---

Built with ❤️ by the GRC Lab at Illinois Institute of Technology