Source and classification
- Source: ZNS Network Service 1.2.0 Layer 4 Load Balancing performance risk record.
- Classification: Conditional performance risk.
- Confirmation basis: Direct stress testing of the backend service can reproduce this behavior without passing through Layer 4 Load Balancing. Therefore, it cannot currently be attributed solely to a Layer 4 Load Balancing logic error.
Issue metadata
- Issue ID: ZCF-4847.
- Component/product: Network Service and the OvsDpdk data plane.
- Status: Not closed.
- Priority/severity: P1.
- Affected version: 1.2.0.
Problem description
Under specific OvsDpdk high-CPS short-connection stress, a small number of connect timeout or read reset events may occur. The same behavior can be reproduced during direct backend service stress testing and is more likely related to the OVS-DPDK/vhost-user TX stress path.
Symptoms
- A small number of connection timeouts or read resets occur during high-CPS short-connection testing.
- The same behavior may be observed both when directly accessing the backend service and when traffic passes through the LB.
- PMD threads may maintain high CPU utilization.
Impact
The success rate may decrease under a specific stress model. This behavior does not mean that ordinary functional traffic will necessarily fail.
Workaround
- When the issue occurs, reduce concurrency and retain the original test parameters and results.
- Retest with the same hardware, VM instances, concurrency, request size, and PMD/RXQ configuration.
- Collect the success rate, CPS,
ovs_tx_failure_drops, vhost counters, host/guest CPU usage, and backend logs at the same time. - Do not determine the service capacity limit only because PMD threads approach 100% CPU utilization. DPDK busy-poll itself keeps CPU utilization high.
Affected products
- Product: ZCF.
- Component: Network Service.
- Version: 1.2.0.
Fix status
- Current status: Analysis continues.
- Related patch/fix package: No separate patch information is available.