Aptos Unpacked: Scaling Beyond Limits

The blockchain industry is rapidly advancing, with an increasing need for systems that can handle the surge in demand for transacting and interacting with on-chain applications. However, scaling blockchain networks while maintaining decentralization and security remains a significant challenge. 

Aptos has made strides in addressing this challenge through its innovative approach to horizontal scaling. In this research piece, we delve into two recent innovations introduced by Aptos for scaling blockchains: Zaptos and Shardines.

Pipelined Architecture

Aptos’ pipelined blockchain architecture contributes to scaling the network's overall throughput. A submitted transaction goes through four major stages:

1. Consensus: Handles payload dissemination and transaction ordering.
2. Execution: Executes transactions in a block.
3. Certification: Certifies the validity of newly executed blocks.
4. Commit: Commits certified blocks to storage.

Aptos’ pipelined blockchain architecture enables different processing stages of separate blocks to overlap, thereby maximizing resource utilization through parallelization and improving overall blockchain throughput.

Zaptos

Aptos’ introduction of Zaptos features three key optimizations aimed at further reducing end-to-end transaction latency in its pipelined blockchain architecture, which is the time required for submitted transactions to get through all four stages of processing.

Optimistic Execution

Instead of waiting for transaction ordering to be finalized among validators, Zaptos allows transactions in the proposed block to be speculatively executed as soon as the previous block has been executed. Optimistic execution reduces latency and accelerates transaction processing.

Optimistic Commit

Like optimistic execution, optimistic commit reduces latency by allowing blocks from optimistic execution to be committed to storage before full certification. If the committed states do not match the transaction order agreed upon from consensus, they are reverted from storage. 

Certification Optimization

The third optimization enables validators to commence the certification process of an optimistically executed block earlier rather than waiting for transactions to be formally ordered. This allows the certification stage to run in parallel with the final round of consensus, further reducing latency. 

By implementing these optimizations, Zaptos significantly lowers pipeline latency while maximizing resource utilization to achieve high throughput. In a testing environment, Zaptos reduces latency by 40% under a load of 20,000 transactions per second (tps).

Modular Sharding

Sharding is a commonly discussed technique in blockchain scaling. It involves dividing a blockchain’s data into smaller segments called “shards,” each of which can be processed in parallel. This reduces the load on individual nodes, allowing for faster transaction processing and significantly increasing the overall network throughput.

While theoretically promising, sharding presents several inherent challenges. The most significant issues include the complexity of managing the communication and coordination between shards while ensuring that the network remains secure and consistent across different shards. 

Aptos has developed a modular approach to scaling by dividing its blockchain into three core layers—storage, consensus, and execution. Rather than implementing sharding to the entire blockchain architecture as a whole, Aptos decouples each of these components and optimizes them independently.

By breaking down a complex system into smaller modules, Aptos can implement scaling on each layer more easily.

Storage Layer

The storage layer is responsible for holding data associated with the blockchain in a RocksDB database, including smart contracts and states. These data are stored in the form of Jellyfish Merkle Tree (JMT), which is a variant of sparse Merkle tree designed to facilitate sharding and ensure an even distribution of the data load across various shards.

Consensus Layer

The consensus layer runs the blockchain’s consensus engine, bringing all nodes into agreement on a single network state. Aptos implements a Narwhal-based Quorum Store, which decouples data dissemination from transaction ordering. This approach allows validators to disseminate data asynchronously in parallel, reducing latency caused by the leader bottleneck. 

Execution Layer

The execution layer executes transactions using the Block-STM engine. Implementing sharding in an execution environment faces several key challenges, including conflict handling and latency arising from cross-shard communication.

Conflicts can occur when two transactions from different shards attempt to modify the same state concurrently (i.e., write-write conflicts) or when a transaction reads a state that is subsequently modified by another transaction (i.e., write-read conflicts).

Shardines

Aptos introduces Shardines, a novel sharded execution engine that features several key innovations designed to achieve high throughput, reduce cross-shard communication latency, and efficiently handle conflicts. 

Hypergraph Partitioning Algorithm

Shardines employs a hypergraph partitioning algorithm to minimize the amount of necessary cross-shard communication and its associated wait times. This algorithm analyzes conflicts between transactions and organizes them in a way that reduces the amount of resource transmission between nodes, ultimately decreasing the need for time-consuming data transfers between shards.

In other words, by partitioning the transactions more intelligently, Aptos optimizes resource usage and reduces the delays that typically plague cross-shard communication.

Shards under Shardines are dynamic, meaning the shards from which resources are accessed vary from block to block, adapting to the dynamic environment of transaction execution.

Micro-batching & Pipelining

Another feature of Shardines is its novel micro-batching and pipelining strategy. Transactions within a block are split into smaller batches, and these transaction batches are then pipelined through various execution stages: 

1. Transmit a transaction batch from the coordinator node to a shard node.
2. Fetch the data and dependencies for the transactions in the batch.
3. Execute transactions in the batch.
4. Transmit the execution results to other shards containing dependent transactions.
5. Transmit the execution results to the coordinator.

Mirco-batching helps break down large data packets, reducing latency associated with serializing and deserializing large chunks of data. Combined with pipelining, this approach enables efficient resource utilization, allowing transaction batches to be executed (stage 3 in Figure 6) without waiting for all other batches to arrive (stage 1). 

In other words, this method achieves parallel execution and high throughput by enabling different execution stages of different transaction batches to overlap.

Delta Aggregation

Delta aggregation is another critical feature of Shardines. When multiple transactions modify shared resources (i.e., write-write conflicts), such as minting events affecting the total token supply, Shardines computes the state change (i.e., delta) for each transaction during execution in parallel and aggregates the results at the end of the execution stages.

By avoiding the need for sequential execution even when encountering conflicting transactions, Shardines ensures the efficient handling of resources affected by multiple transactions without creating bottlenecks. 

These features enable Aptos to achieve near-linear throughput scaling with increasing shard count, reaching over 1 million tps with 30 shards in a testing environment with non-conflicting transactions.

Even in scenarios with conflicting transactions, its throughput still exceeds 500,000 tps, demonstrating its resilience under more realistic conditions where many blockchain users interact with the same on-chain resources. This solidifies Aptos’ position as a pioneer of parallelism in the blockchain industry.

Takeaways

In summary, Aptos has developed two groundbreaking technologies—Zaptos and Shardines—that address the challenges of scaling blockchain networks.

• Zaptos reduces end-to-end transaction latency by implementing parallelization and optimistic operation strategies within Aptos’ pipelined blockchain architecture. 
• Shardines enhances the execution layer through intelligent partitioning, pipelining, and efficient conflict resolution, dramatically boosting throughput while minimizing cross-shard communication.

Together, these innovations enable Aptos to scale efficiently, positioning it at the forefront of blockchain parallelism. By achieving high throughput and low latency, Aptos paves the way for the widespread adoption of decentralized applications and services capable of handling large transaction volumes with minimal delays.