Skip to main content

Hybrid Infrastructure Benchmark

4 min read
AI documentation guide

This test plan compares network, file-storage, and image-distribution paths between cloud and on-premises sites. It contains no execution results or performance rankings. Record hardware and software versions, traffic paths, load, repetitions, and raw results for each test.

Network Performance​

EKS Hybrid Nodes Network​

Testing Planned

This benchmark is currently being prepared.

Metrics

  • Latency and packet loss between cloud and on-premises sites
  • Observed throughput and available circuit capacity over the VPN or Direct Connect path
  • Communication between AWS-hosted and on-premises Pods, with node and Pod addresses, routing, and CNI conditions fixed

This is a cross-site comparison. A test between two AWS Regions needs a separate configuration and result set.

SR-IOV Network Acceleration​

Metrics

  • Throughput on the same physical NIC using an assigned SR-IOV VF versus a specified virtual-switch and virtual-NIC path
  • RDMA and GPUDirect RDMA operation and performance in a separate environment with supported GPUs, NICs, drivers, and fabric
  • Absolute latency and its difference under the same load

SR-IOV itself exposes virtual functions, so “virtual NIC” alone does not identify the baseline. Establish the required hardware and transport path before an RDMA test; connecting VPN or Direct Connect does not establish RDMA capability. Calculate improvement percentages only when a defined baseline and measurements exist.

Storage Performance​

Hybrid File Storage​

Metrics

  • Sequential and random read/write IOPS and throughput with fixed block size, queue depth, and cache conditions
  • Large-file transfer speed by file size and path
  • Shared-filesystem performance by client count and access pattern

Container Registry​

Harbor Registry Performance​

Metrics

  • Pull time for the same image digest within and across sites, separating empty caches from cached layers
  • Replication delay from completed push until the same digest can be resolved and pulled at the destination registry
  • Throughput and failure rate by concurrent pull count