Paper summary
CPU allocation in isolated Fabric
With the right CPU allocation, isolated Hyperledger Fabric containers came close to native performance in the tested setup.
Hyperledger Fabric was run in native containers and in Kata Containers, which put workloads in lightweight virtual machines. The question was how much transaction throughput and latency change under that deployment, and how the allocation of CPU between Fabric components changes those metrics. The orderer sorts transactions; the peer verifies transactions and runs chaincode.
Native and Kata configurations
The benchmark sends 300 transactions per second and measures transaction throughput, average latency and CPU use. It tests the effect of different vCPU allocations to the Kata orderer and peer with four, six and eight physical cores, alongside a native-container comparison.
Throughput and latency on four cores
On four physical cores, the best tested Kata setting assigned one vCPU to the orderer and two to the peer. Its throughput and latency differed from the native optimum by 2.15% and 2.56%, respectively.
Two-core comparison
With two physical cores and a uniform allocation, Kata delivered 38.02% lower throughput and 89.99% higher average latency than native containers. CPU use was 41.35% higher in the same comparison.