Hello corelibs-dev and build-dev, Back in May, Brian Croom implemented performance testing in swift-corelibs-xctest: https://github.com/apple/swift-corelibs-xctest/pull/109
I’d love to see Swift developers use this feature to measure the performance of their code. I think we’ll need to add functionality to swift-corelibs-xctest and SwiftPM in order to do so. The problem: recording performance test baselines In order for performance tests to be useful, Apple’s Xcode provides a way to record “baseline” metrics. Baseline metrics allow a developer to indicate “this performance test should never be slower than 1.2 seconds on average, with 10% standard deviation as ‘wiggle room’”. When Apple XCTest tests are run, they are informed of the baseline metrics that have been set in Xcode. Apple XCTest performance tests that have a baseline registered will fail if performance becomes slower than the acceptable amount. If we could provide swift-corelibs-xctest with a mapping from each performance test to its baseline metric, it would be easy to write the code to fail a test if it didn’t perform well enough. That mapping, however, is the tricky part. Here’s why: - The mapping needs to group metrics based on the host machine running the test. Performance will of course vary based on the hardware, so it’s important to make sure performance baselines set on a Raspberry Pi aren’t used when testing on a Mac Pro. - The mapping also needs to group metrics based on the target machine. Using Apple XCTest, a developer can start a test suite run from their MacBook Pro (macOS 64-bit), and see the results of the performance tests when run on their iPhone 6s (iOS armv7s). I don’t think this is relevant to swift-corelibs-xctest just yet — as far as I know, SwiftPM is not capable of cross-compilation, so the host machine will always be identical to the target machine. Still, we should design something flexible enough for this scenario. Xcode’s solution: plist files Xcode’s solves this problem using two kinds of .plist files. I tried creating a sample project, named Perforate.xcodeproj, which contained a single performance test. Here’s what Xcode created: <!-- Perforate.xcodeproj/xcshareddata/xcbaselines/DA77262F1D447DB300735C93.xcbaseline/Info.plist --> <?xml version="1.0" encoding="UTF-8"?><!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd";><plist version="1.0"><dict> <!-- runDestinationsByUUID: These are the host/target machine groups. --> <key>runDestinationsByUUID</key> <dict> <!-- It appears each group is given a UUID, but to be honest, I'm not sure why. It seems like these should be "keyed" on aspects of the host/target machines. As-is, I imagine Xcode and Apple XCTest need to traverse each group's `localComputer`, `targetArchitecture`, and `targetDevice`'s values in order to find a match. --> <key>8CE9E051-9AB6-44AF-8B80-F2DEFD409CB5</key> <dict> <!-- Information about the host machine: number of CPUs, cores, etc. --> <key>localComputer</key> <dict> <key>busSpeedInMHz</key> <integer>100</integer> <key>cpuCount</key> <integer>1</integer> <key>cpuKind</key> <string>Intel Core i7</string> <key>cpuSpeedInMHz</key> <integer>2800</integer> <key>logicalCPUCoresPerPackage</key> <integer>8</integer> <key>modelCode</key> <string>MacBookPro11,3</string> <key>physicalCPUCoresPerPackage</key> <integer>4</integer> <key>platformIdentifier</key> <string>com.apple.platform.macosx</string> </dict> <!-- The target architecture and device are stored as separate keys. --> <key>targetArchitecture</key> <string>x86_64</string> <key>targetDevice</key> <dict> <key>modelCode</key> <string>iPhone8,2</string> <key>platformIdentifier</key> <string>com.apple.platform.iphonesimulator</string> </dict> </dict> </dict></dict></plist> <!-- Perforate.xcodeproj/xcshareddata/xcbaselines/DA77262F1D447DB300735C93.xcbaseline/8CE9E051-9AB6-44AF-8B80-F2DEFD409CB5.plist --> <!-- Notice that this file is named after the `runDestinationsByUUID` key from the first file: 8CE9E051-9AB6-44AF-8B80-F2DEFD409CB5. --> <?xml version="1.0" encoding="UTF-8"?><!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd";><plist version="1.0"><dict> <key>classNames</key> <dict> <key>PerforateTests</key> <dict> <!-- The metrics are mapped by class name and test method name to performance metrics. --> <key>test_uniqueOrdered_performance</key> <dict> <!-- There are several categories of performance metrics. The only one publicly available in Apple XCTest so far is wall clock time. --> <key>com.apple.XCTPerformanceMetric_WallClockTime</key> <dict> <key>baselineAverage</key> <real>0.5</real> <key>baselineIntegrationDisplayName</key> <string>Local Baseline</string> </dict> </dict> </dict> </dict></dict></plist> Proposed solution for SwiftPM/swift-corelibs-xctest: JSON files I think we can mimic Xcode’s approach here. Here’s what I’m proposing: - swift-corelibs-xctest’s test runner should take a --performance-metrics <PATH> argument, where <PATH> is the location of a file containing JSON that looks pretty much exactly like the 8CE9E051-9AB6-44AF-8B80-F2DEFD409CB5.plist from above: { "classNames": { "PerforateTests": { "test_uniqueOrdered_performance": { "baselineAverage": "0.5", "baselineIntegrationDisplayName": "Local Baseline" } } }} - SwiftPM’s swift test command should also take a --performance-metrics <PATH> argument, where <PATH> is the location of a file containing JSON that looks pretty much exactly like the xcbaselines/DA77262F1D447DB300735C93.xcbaseline/Info.plist from above (by default, --performance-metrics could be set to the same path as the swift test --build-path directory): { "runDestinationsByUUID": { "8CE9E051-9AB6-44AF-8B80-F2DEFD409CB5": { "localComputer": { "busSpeedInMHz": "100", # ... }, "targetArchitecture": "x86_64", "targetDevice": { # We might need to change these keys, since "modelCode" seems very Apple-specific. "modelCode": "linux", "platformIdentifier": "Ubuntu 15.04", } } } } Personally, I think the format of the plist files Xcode and Apple XCTest generate could be improved. Still, I think it’d be nice to stick to the same format (as much as possible) for swift-corelibs-xctest, just to keep things simple. Thoughts? I admit that I don’t have much experience using Apple XCTest’s performance testing functionality, so I might be missing something here. Does anyone have any feedback on this idea? I’d like to incorporate your feedback, and perhaps submit a Swift Evolution proposal for this feature. - Brian Gesiak
_______________________________________________ swift-corelibs-dev mailing list swift-corelibs-dev@swift.org https://lists.swift.org/mailman/listinfo/swift-corelibs-dev
