Sharding Database with RDMA

Blockchain systems face significant scalability challenges due to ever-growing state sizes. To address this, InfiniSVM implements a sharded key-value store architecture, integrating Remote Direct Memory Access (RDMA) to enable ultra-low latency cross-node data access.

​

Sharded State Management Architecture

The InfiniSVM database consists of sharded state storage across distributed nodes. Unlike Solana’s 10MB account size limit, this system supports arbitrary account sizes, dynamically optimizing data distribution.

​

Database Node Components

Each database node consists of three core components:

1. Jump Table (Memory Indexing)

   • Maps 32-byte account addresses to their memory locations, data length, and version numbers.
   • Ensures fast lookups for state access with minimal CPU overhead.

2. Contiguous Memory Storage

   • Stores account state data along with synchronization primitives.
   • Efficient for most transactions, as state size rarely changes.
   • If state resizing occurs, data is redistributed during rebalancing to maintain contiguous storage.

3. Local Cache for Frequently Accessed Data

   • Implements an LRU (Least Recently Used) eviction policy.
   • Reduces cross-node fetches by keeping hot accounts in memory.

​

RDMA-Optimized Cross-Node Data Access

To enable efficient cross-node execution, SOLAYER CHAIN utilizes InfiniBand RDMA, providing:

• Ultra-low latency (nanosecond-level) memory access across nodes.
• Zero-copy data movement, bypassing the operating system overhead.
• Reduced CPU load, offloading memory transfers to specialized network hardware.
• Consistent transaction processing speeds, even when data is fetched remotely.

By eliminating traditional network bottlenecks, RDMA ensures high-throughput, low-latency execution across sharded storage nodes.

​

Dynamic Load Balancing & Rebalancing Mechanism

To maintain optimal performance, InfiniSVM employs a background rebalancing mechanism that:

Monitors Account Access Patterns

• Analyzes historical memory access footprints to identify high-traffic accounts.

Optimizes Data Locality

• Frequently co-accessed accounts are migrated to the same node.
• Minimizes remote memory fetches and improves execution efficiency.

Adaptive Data Redistribution

• Rebalances data based on:
  • Access frequency
  • Account size
  • Network topology
  • Node capacity and current load

This adaptive approach ensures even workload distribution, reducing cross-node dependencies while maximizing throughput and parallel execution efficiency.

By combining sharded key-value storage with RDMA-powered cross-node memory access, InfiniSVM achieves:

• Scalable, high-performance blockchain state management.
• Reduced transaction latency via direct memory access.
• Dynamic workload balancing for optimized execution efficiency.

This architecture ensures sustainable scalability, making it well-suited for high-throughput, decentralized applications.