performance analysis

Bilingual

The mean means nothing: data visualization to debug a latency problem
2 months ago
- Performance improvements in benchmarks may not reflect live production due to noise in metrics.
- Multiple statistics (mean, median, p95, p99) can tell contradictory stories about the same data change.
- Cumulative distribution functions (CDFs) visualize the entire distribution, revealing shifts and crossing points.
- Rollout data over time shows emerging regression patterns that aggregates hide.
- Bimodality in latency often has a root cause, such as cache hits vs. misses based on response size.
- Jointplots can isolate populations by variables like response size to identify actionable causes.
- Multiple visualizations are needed to fully understand data instead of relying on a single number.
How did Apple cut launch time by 30% in iOS 27?
2 months ago
- Launch time is a critical metric; even a 1% improvement saves significant time across billions of daily launches.
- Apple claimed a 30% boost in app launch times in iOS 27, achieved through an optimized CPU scheduler and preloading key app data.
- The article explains launch phases: cold launch (process doesn't exist), warm launch (cached data), and no 'hot launch' (just resume).
- Using Instruments, the author measured iOS 26 vs iOS 27: cold launches were 21% faster, warm launches 23% faster, with most gains in pre-main phases.
- Improvements in iOS 27 include faster closure rebuilding (1.5x), fixups (2.5x), and static initializers (3x), while I/O steps remained similar.
- Apple has historically improved launch times (e.g., dyld3, CPU clock warming), but actual gains are often modest, emphasizing the need for developers to optimize their own launch paths.
Understanding the Go Runtime: Profiling
2 months ago
- The reflect package uses type descriptors compiled at build time, while profiling captures runtime data through call stacks.
- Go provides five profiles (CPU, heap, block, mutex, goroutine) with a shared pprof format based on a simple protocol buffer structure.
- Profiles share a common skeleton with indirection: samples reference locations, which reference functions, which reference strings in a shared table for efficiency.
- CPU profiling uses asynchronous signal-driven collection with a ring buffer to avoid allocation and locking in handlers.
- Heap, block, and mutex profiles use in-place aggregation into hash tables (buckets) keyed by call stacks, triggered by specific runtime events.
- Heap profiling samples allocations based on a randomized rate, scaling counts to estimate totals, and tracks live vs. allocated objects.
- Block and mutex profiles record contention events, with selective sampling by duration or random fraction, and scale values to estimate true totals.
- Goroutine profiling takes an on-demand snapshot of all live goroutines, using a cooperative approach to avoid long stop-the-world pauses.
- All profiles ultimately produce the same pprof file format, with differences in how and when stacks are captured and aggregated.
HammerDB v6.0 Response Times, Percentiles and Reservoir Sampling
2 months ago
- A key mistake in database benchmarking is oversizing workloads to the point where response times reach hundreds of milliseconds or seconds, making throughput numbers misleading if latency is high.
- HammerDB v6.0 introduces new response time metrics, including full percentile reporting and box plots, to visualize latency profiles for complex transactions, moving beyond averages to show medians, spread, and outliers.
- The update adds reservoir sampling for long runs, ensuring practical response time analysis without data bottlenecks, even with large transaction volumes.
- It emphasizes reviewing throughput and latency together; if throughput rises while latency spikes to hundreds of milliseconds, it indicates workload overload, and HammerDB v6.0 makes this easier to detect.

performance analysis

The mean means nothing: data visualization to debug a latency problem

How did Apple cut launch time by 30% in iOS 27?

Understanding the Go Runtime: Profiling

HammerDB v6.0 Response Times, Percentiles and Reservoir Sampling