Scale or Die at Accelerate 2025: Scale to win: agave's performance arc

In a groundbreaking presentation at Accelerate 2025, Alessandro Decina of Anza's performance team unveils revolutionary improvements to Solana's scalability, challenging the notion of theoretical limits and showcasing concrete solutions that promise to dramatically increase the network's throughput.

Summary

Alessandro Decina, a key member of Anza's performance team, delivered a compelling presentation at Accelerate 2025, addressing misconceptions about blockchain scalability and revealing significant performance improvements for Solana. Decina argues that perceived scalability limits are often just bugs waiting to be fixed, and demonstrates this through in-depth analysis of Solana's performance bottlenecks.

The presentation focuses on two major areas of improvement: the Turbine protocol and epoch boundary processing. By implementing an XDP-based version of Turbine, Decina shows how Solana can handle orders of magnitude more network traffic than it currently does on mainnet. Additionally, he highlights a critical bug fix for epoch boundary processing that could eliminate block skipping, a persistent issue in the network.

These improvements are not just theoretical – they represent tangible solutions that are either already implemented or on the verge of deployment. Decina's work demonstrates that Solana is poised for significant scaling, potentially enabling larger transactions and doubling block space without compromising performance.

Key Points:

Debunking Scalability Myths

Decina begins by challenging the common belief that blockchain networks have reached an imaginary limit of scalability. He argues that what many perceive as limits are actually bugs that can be fixed. This perspective shift is crucial, as it moves the conversation from theoretical constraints to practical solutions.

Decina cites the recent 1 million TPS demo as evidence that Solana's virtual machine is capable of processing transactions at a much higher rate than currently seen on mainnet. He emphasizes that the bottlenecks lie not in the core processing capabilities, but in surrounding systems like transaction ingestion and the Turbine protocol for block propagation.

Turbine Protocol Improvements

A significant portion of the presentation focuses on improvements to the Turbine protocol, which is responsible for propagating blocks across the network. Decina demonstrates how the current implementation struggles under high load, particularly when simulating a cluster of 5,000 peers.

Using detailed profiling, he shows that the current Turbine implementation spends a large amount of time on inefficient operations, including unnecessary kernel calls and outdated mitigation techniques. To address this, Decina introduces an XDP (eXpress Data Path) based implementation of Turbine.

This new implementation dramatically reduces CPU usage, allowing a single core to handle the work that previously required 12 cores. More importantly, it enables the processing of 2.3 million packets per second – orders of magnitude more than what Solana currently handles on mainnet. Decina has tested this implementation up to 100 gigabits per second and is eager to push it even further.

Epoch Boundary Optimization

Another critical improvement Decina discusses is the optimization of epoch boundary processing. Currently, Solana experiences a consistent skip in block production at epoch boundaries, leading to a network-wide skip rate of about 0.5%.

Decina's analysis reveals that this issue stems from inefficient parallelization in the implementation of partition rewards. By addressing this bug, Solana can potentially eliminate block skipping at epoch boundaries, solving one of the last remaining causes of skipped blocks in the network.

Collaborative Development

Decina concludes by encouraging community involvement in performance improvements. He invites developers to reach out and contribute to ongoing optimization efforts, emphasizing the collaborative nature of Solana's development.

Facts + Figures

• Anza now has five full-time members working on Solana's performance team
• The current Solana network has a skip rate of about 0.5%, down from as high as 30% during congestion periods last year
• Decina's XDP-based Turbine implementation can process 2.3 million packets per second
• The new Turbine implementation reduces CPU usage from 12 cores to just one, with capacity to spare
• Decina has tested the new Turbine implementation up to 100 gigabits per second
• The current Turbine implementation struggles with a simulated cluster of 5,000 peers
• Epoch boundary processing currently causes a block production pause of nearly two seconds
• The inefficient parallelization at epoch boundaries results in only 4.5% of time spent on actual work, with the rest lost to contention

Top quotes

1. "We just have bugs, and bugs are easy to fix. Once we find the bugs, we fix them, and we get so much faster."
2. "My personal goal, I want one thing in life, and I want to double block space, and I want larger transactions."
3. "We can send so many more packets, right? In this test, we were saturating 25 gigabits per second, but I have tested up to 100 gigabits per second."
4. "We shouldn't be skipping any blocks, right?"
5. "Find me, and I will find you stuff to do, and then you send me PRs, and everyone is happy."

Questions Answered

What is the main argument Alessandro Decina makes about blockchain scalability?

Decina argues that perceived scalability limits in blockchain networks are often just bugs waiting to be fixed, rather than fundamental technological barriers. He demonstrates this by showcasing significant performance improvements in Solana's Turbine protocol and epoch boundary processing, which can dramatically increase the network's throughput and reliability.

How does the new XDP-based Turbine implementation improve Solana's performance?

The XDP-based Turbine implementation dramatically reduces CPU usage, allowing a single core to handle the work that previously required 12 cores. It enables processing of 2.3 million packets per second, which is orders of magnitude more than what Solana currently handles on mainnet. This improvement allows for much higher network throughput and more efficient block propagation.

What is causing the block skipping issue at epoch boundaries, and how is it being addressed?

The block skipping at epoch boundaries is caused by inefficient parallelization in the implementation of partition rewards. The current implementation creates excessive threads, leading to contention and slow execution. By fixing this bug, Solana can potentially eliminate block skipping at epoch boundaries, solving one of the last remaining causes of skipped blocks in the network.

How significant are the performance improvements Decina presents?

The improvements are substantial. The new Turbine implementation can handle orders of magnitude more network traffic than the current mainnet, potentially enabling larger transactions and doubling block space. The epoch boundary optimization could eliminate the 0.5% skip rate currently experienced by the network. These improvements represent a significant leap in Solana's scalability and reliability.

How can developers contribute to Solana's performance improvements?

Decina encourages developers to reach out to him directly if they're interested in contributing to Solana's performance improvements. He offers to find tasks for interested developers and welcomes pull requests, emphasizing the collaborative nature of Solana's development process.