Storage Benchmark Proposal Edgar StPierre, EMC

Proposal Project Name: STORPERF Repo Name: STORPERF Category: Requirements Project Lead: Edgar StPierre, EMC Project Goal: Provide tools to measure block and object storage performance in an NFVI. Ideally, this is complemented with an effort to characterize typical VF storage performance requirements.

Use Cases (Scope) 1.Characterize expected storage performance behavior of an NFVI for any type of block or object storage deployment in an OPNFV lab 2.Troubleshoot actual storage performance in a production NFVI (obviously, to be used with caution) 3.Workload analysis during NFVI staging before deployment. (Integrate with project Bottlenecks.)

Project Deliverables Brahmaputra: – Definition of Performance Test Cases 1.Block – multiple block sizes with fixed queue depths and target data sizes; measure both read and write performance 2.Object – both small and very large target data size to include video streaming emulation – Definition of basic metrics to measure performance: Max IOPS under various loads, Average I/O Latency, more? – Definition of test process Including relative applicability of test processes to different VNF workloads Including robustness testing for impaired storage environments – On track to deliver benchmark test tools in C release Identify open source tool(s) such as FIO, IOmeter, VDBench, The Grinder, Locust, and/or JMeter Identify integration points with Qtip, Yardstick, and/or Jenkins tool chain

Block Performance Test Cases Storage capacity: min specified, max TBD Preconditioning of storage – Storage performance degrades until it reaches steady state[1] – Period TBD, but est 2-6 hours Test queue depths of 1, 16, & 128 Test block sizes of 4KB, 8KB, 64KB, 1MB Test 5 workloads: 4 corners and 1 mixed Metrics – IOPS report measured to a max latency? (or when it hits “the wall”?) for each workload – Avg Latency report for each workload random read write sequential Workloads

Object Performance Test Cases Swift API (?) Measure max concurrency with smaller data size (GET/PUT) Measure max TPS using variable object size payloads at max concurrency (1KB, 10KB, 100KB, 1MB, 10MB, 100MB, 200MB) 5 different GET/PUT workloads for each: 100/0, 90/10, 50/50, 10/90, 0/100 Separate metadata concurrency test using List & HEAD Metrics – Transactions/second – Error rate – Per-test average latency

Future Project Deliverables Future test extensions – Expand captured performance metrics (e.g., I/O Latency variation for object streaming) – Integration with Qtip for automated reporting – Integration with Yardstick for automated execution Separate deliverable capturing corresponding typical VF storage performance requirements using the same metrics, for VFs that require block or object storage IO Captured through collaborative polling of VF producers, preferably using empirical data Used to drive pass/fail criteria for measurements

Application of Test Tool Executes as VM(s) in test environment – Manual deployment, or automated in tool chain Possible Target SUT: – Direct attached block storage (local LUNs) – External or distributed block storage (iSCSI) – External or distributed object storage (HTTP) Host 1 DAS Storage Benchmark VM Host 2 External Block Storage Storage Benchmark VM Host 3 External Object Store Storage Benchmark VM Local LUN iSCSIHTTP

Reporting Benchmark Tool will produce reports of SUT performance for defined test cases – Accessed directly – Accessed via Qtip – Accessed via Yardstick

Automation / Testability Full automation via integration with Yardstick and Qtip tool chain for testing of NFVI test environments

Contributors Edgar StPierre, EMC Chanchal Chatterjee, EMC Iben Rodriguez, Spirent Jose Lausuch, Ericsson Ferenc Farkas, Ericsson Vishal Murgai, Cavium Networks (add to etherpad)

References [1] For example, see and