Project: StorPerf - Storage Performance Benchmarking for NFVI
Open Bug List
|STORPERF-56||Cannot delete stack if create failed||Jun 16, 2016||May 02, 2017||Unassigned||Open||Unresolved||Danube 3.0|
Key Project Facts
StorPerf Project Scope
StorPerf testing addresses both block storage and object stores, though using different test suites. There is limited value in testing locally attached storage, so this is primarily about testing distributed/external storage environments.
StorPerf is intended to run standalone test benchmark tools, plus provide integration with test frameworks such as Qtip and Yardstick.
There are three applicable use cases for these storage performance benchmarks:
- An OPNFV test lab manager wants to characterize expected storage behavior in a test NFVI deployment. This will include both a preconditioning phase for each storage environment as well as the broadest set of test cases across all identified storage services. This will provide VNF test applications with information about expected storage performance. This will integrate with existing test lab tool chains.
- A Service Provider wants to validate storage performance in an NFVI staging environment prior to production deployment. This will validate expected performance expectations using pass/fail conditions using the same preconditioning and test cases as for a test lab. This will integrate with project Bootstrap.
- A Service Provider wants to isolate performance problems in a production NFVI environment. This will use a much narrower set of test cases to minimize impact on the production environment. This will utilize a manual deployment and control of the test VMs.
The high level plan for StorPerf is to deliver (minimally) test requirements and test process specifications in the Brahmaputra release timeframe. Block performance testing will lead object testing, and could also be delivered in Brahmaputra, though any such delivery would be asynchronous to, and largely independent of, the Brahmaputra release mechanism. In the C release, we will complete object store testing and integration with Qtip and Yardstick.
Project Planning: TBD
This is an outline of test cases. A specification will be written capturing actual tests and steps. And of course, the input to the test process will be determined by community participation.
Assuming iSCSI-attached storage, though local direct attached storage, or Fibre Channel-attached storage could also be tested.
- Preconditioning of defined Logical Block Address range (period TBD)
- Testing across each combination of: Queue Depths (1, 16, 128) and Block sizes (4KB, 64KB, 1MB)
- For each of 5 workloads: Four corners (100% Read/Seq, Write/Seq, Read/Random, Write/Random) and mixed (70% Read/Random).
Assuming an HTTP-based API, such as Swift for accessing object storage.
- Determine max concurrency of SUT with smaller data size (GET/PUT) tests by determining performance plateau
- Determine max TPS of SUT using variable block size payloads (1KB, 10KB, 100KB, 1MB, 10MB, 100MB, 200MB)
- Use 5 different GET/PUT workloads for each: 100/0, 90/10, 50/50, 10/90, 0/100
- Perform separate metadata concurrency test for SUT using List and Head operations
Especially looking for workload recommendations for testing in this area.
Initially, metrics will be for reporting only and there will not by any pass/fail criteria. In a future iteration, we may add pass/fail criteria for use cases which are testing viability for known workload requirements.
Block Storage Metrics
The mainstays for measuring performance in block storage are fairly well established in the storage community, with the minimum being IOPS and Latency. These will be produced in report/tabular format capturing each test combination for:
- IOPS at a fixed max latency (TBD; we could also choose to report IOPS when the test hits the latency "wall"). Note that throughput data can be calculated based on IOPS * block size.
- Avg Latency for each workload at different IOPS levels
Object Storage Metrics
Object storage delivers different storage characteristics than block storage, and so the metrics used to charaterize it vary to some degree:
- Transactions per second (throughput can also be calculated from TPS * object size)
- Error rate
- Per-test average latency
See also future extensions below.
Future Project Extensions
These are 2nd+ release ideas for extending StorPerf.
- Definition of more extensive metrics to measure performance (e.g., I/O Latency variation for object streaming); some of these may require contributions to upstream open source test tools
- Time-to-first-write for newly provisioned block volumes. This is intended to measure the impact of zero-out functions performed by storage systems when a volume is provisioned.
- Full integration with Qtip and Jenkins for automated deployment and reporting
- Create separate deliverable (document) to capture typical/expected VF storage performance requirements using the same metrics, for those VFs that require block or object storage I/O. This can be used to define pass/fail criteria for test lab deployments.
Daniel Smith (lmcdasm)
Edgar StPierre (estpierre)
Iben Rodriguez (ibenr)
Mark Beierl (mbeierl)
qi liang (QiLiang)
Tim RAULT (trault14)
- Chanchal Chatterjee, EMC firstname.lastname@example.org
- Vishal Murgai, Cavium Networks Vishal.Murgai@caviumnetworks.com
- Vikram Dahm, Dell V_Dham@Dell.com
- Stephen Blinick email@example.com
- srinivas tadepalli, Tata Consultancy Services
- Nataraj Goud, Tata Consultancy Services firstname.lastname@example.org
- Dennis Qin, EMC
- Sam Decker, Unaffiliated, Algonquin College Student
Aric Gardner (agardner)
Edgar StPierre (estpierre)
Jose Lausuch (jose.lausuch)
Mark Beierl (mbeierl)