Server-Side Hardware Impact on Throughput: E2 vs. C4D

An in-depth benchmark analysis demonstrating hardware impact on total throughput (Intel Broadwell vs. AMD Turin).

Introduction

RogueDB throughput is bounded by two primary elements: database engine execution efficiency and client-side request serialization. Profiling shows that serialization overhead and bidirectional gRPC stream management place intensive demands on CPU cycles. The following demonstrates and supports our decision for transparency in the underlying hardware generation allocated to customer database instances.

For this benchmark, the client hardware uses an AMD Turin with 4 cores. Only the server side hardware changes, and testing procedures remain identical between the two.

Setup

To measure the performance delta introduced by server compute hardware, we executed the YCSB General Purpose benchmark framework. The client hardware remained static while the server compute alternated between Google Cloud E2 and C4D machine classes.

The E2 instance utilizes the Intel Broadwell architecture, while the C4D instance utilizes AMD Turin architecture. Core count was normalized to 4 vCPUs across both setups (16GB RAM for E2; 15GB RAM for C4D). The client server was hosted on a separate 4-core, 15GB C4D virtual machine. Standalone Read operations were evaluated for both sequential and batch operations to assess if payload batching mitigates hardware architecture differences.

Throughput Results

Key Findings

3.1x to 3.6x Throughput Multiplier: Upgrading hardware to AMD Turin yielded an immediate 210% to 260% increase in executed operations per second, confirming that hardware selection plays a critical factor in database performance.
Diminishing Batching Returns: Compounding operations into batch structures improves throughput on legacy hardware, but performance remains capped.

Discussion

The performance variance between the Intel Broadwell and AMD Turin instances highlights the importance of CPU instruction efficiency and total throughput. Because gRPC payload serialization and HTTP/2 framing operations are CPU-bound tasks, hot loops benefit significantly from the latest advancements.

These validation data establish our production infrastructure guidelines: RogueDB utilizes C4D hardware as the baseline environment standard. Despite a 30% increase in standard host pricing relative to E2 infrastructure, the resulting 3x throughput capacity lowers the net cost-per-operation. Furthermore, C4D instances provide native integration with Google Cloud Balanced Hyperdisks, preventing storage IOPS bottlenecks. Workload migrations to Zen 6 AMD EPYC configurations will follow as the hardware reaches general availability on release.

Conclusion

Database vendors' lack of transparency on underlying hardware generation directly impacts cost of operation to customers. Intel vs. AMD does not sufficiently differentiate. Therefore, high level core counts provides minimal insight into the expected performance impacts. Customers have a right to know what they purchase for the underlying hardware.