AWS Graviton4 vs. GCP Axion

This blog post dives into a head-to-head performance comparison of two leading contenders: AWS Graviton4 (powering AWS r8g instances) and Google Axion (powering GCP Axion instances), both built on the advanced Arm Neoverse-V2 architecture. We'll examine their performance with Valkey 8.0.1, a popular in-memory data store.

The Contenders: AWS Graviton4 and Google Axion

AWS Graviton and Google Axion represent the latest generation of ARM-based server processors from Amazon and Google.  Both leverage the Arm Neoverse-V2 CPU architecture, which is specifically designed for cloud computing, machine learning, and high-performance computing (HPC). These custom chips aim to provide superior performance and energy efficiency compared to traditional x86-based alternatives.

The Benchmark: Valkey 8.0.1

To conduct a meaningful comparison, we chose Valkey 8.0.1, a high-performance, open-source in-memory data structure store. Valkey is a fork of Redis and is widely used for caching, session management, and real-time analytics, making it an excellent workload for testing the raw processing and memory capabilities of these instances. Our benchmark setup was configured to ensure a fair comparison:

Valkey Server Cores: The Valkey server was pinned to cores 2 through 7.

Request Parameters: Each experiment used 100 million parallel requests, 256 clients, and a payload size of 1024 KiB.

Performance Metrics: We focused on two key metrics: Requests Per Second (RPS) for throughput and P99 latency (the 99th percentile) for responsiveness.

Experiment 1: Network Performance

The first experiment tested the performance when the Valkey client and server instances were running on separate hosts within the same cluster network. This scenario highlights the efficiency of the underlying network virtualization and interconnects, which are critical for many distributed workloads.

IRQ Pinning: For this test, we used IRQ pinning on cores 0 and 1. This dedicates specific CPU cores to handling network interrupts, preventing them from interfering with the Valkey server's workload and ensuring a more stable and accurate network performance measurement.

Distributed Application
	AWS r8g Instance Results	GCP Axion Instance Results
SET RPS	925,860	790,020
SET P99 Latency	0.431 ms	0.655 ms
GET RPS	941,802	870,920
GET P99 Latency	0.415 ms	0.543 ms

In this network-bound test, the AWS r8g instances consistently outperformed GCP Axion in both SET and GET operations, with higher throughput and lower P99 latency. This suggests that the AWS Nitro System's networking capabilities, which are tightly integrated with the Graviton4 processor, provide a notable advantage for distributed, network-sensitive applications.

Experiment 2: Same-Host Performance

The second experiment evaluated the raw processing power by running both the Valkey client and server on the same host. This test minimizes network overhead and focuses on CPU and memory performance.

Same-Host Application
	AWS r8g Instance Results:	GCP Axion Instance Results:
SET RPS	1,024,894	894,262
SET P99 Latency	0.407 ms	0.367 ms
GET RPS	1,060,186	942,720
GET P99 Latency	0.359 ms	0.303 ms

Here, the results reveal a more nuanced picture. While AWS r8g instances again delivered higher overall throughput (RPS), the GCP Axion instances demonstrated lower P99 latency for both SET and GET operations. This indicates that while AWS's architecture may be optimized for achieving maximum throughput, Google's design seems to prioritize low-latency performance, which is a key characteristic of Valkey's command execution model.

Conclusion and Analysis

The benchmark results paint a clear picture: for this specific workload, AWS Graviton4-based r8g instances lead in raw throughput, while Google Axion instances excel in latency.

AWS r8g (Graviton4): The higher RPS in both experiments suggests that AWS's implementation is highly optimized for parallel and high-throughput workloads, likely due to a tight integration with the AWS Nitro System.

GCP Axion: The lower P99 latency on the same-host test is a significant indicator. It suggests that Google's Axion processor might have a more efficient core design or cache structure that benefits workloads where low-latency performance is paramount.