CSI Performance Benchmarking & Tuning 2 — Questions and Answers
Question 1: Which metric best indicates whether a system is CPU-bound versus I/O-bound during a benchmark run?
- CPU utilization vs. I/O wait percentage (Correct answer)
- Total memory allocated
- Network packet loss rate
- Disk partition count
Correct answer: CPU utilization vs. I/O wait percentage
Comparing CPU utilization to I/O wait time reveals whether the bottleneck is processing power or storage throughput.
Question 2: A load test shows response times increase linearly with user count up to 100 users, then spike sharply. What does this pattern most likely indicate?
- A resource saturation point has been reached (Correct answer)
- Network latency is increasing uniformly
- The application has a memory leak
- DNS resolution is failing under load
Correct answer: A resource saturation point has been reached
A sudden spike after linear scaling indicates a saturation point where a bottlenecked resource can no longer queue additional requests efficiently.
Question 3: What is the primary purpose of a 'warm-up' phase at the start of a performance benchmark?
- Allow JIT compilation and caches to reach steady state (Correct answer)
- Increase concurrent user count gradually
- Reset database indexes before testing
- Validate that all test scripts are syntactically correct
Correct answer: Allow JIT compilation and caches to reach steady state
A warm-up phase lets the JVM JIT-compile hot paths and populate caches so benchmark measurements reflect steady-state performance rather than cold-start overhead.
Question 4: In capacity planning, the term 'headroom' refers to:
- The unused resource capacity reserved for traffic spikes (Correct answer)
- The maximum number of simultaneous connections supported
- The difference between peak and average throughput
- The time required to provision additional servers
Correct answer: The unused resource capacity reserved for traffic spikes
Headroom is the intentional buffer of unused capacity kept available so unexpected load spikes do not immediately saturate the system.
Question 5: Which profiling technique adds the least overhead and is most suitable for production environments?
- Sampling profiler (Correct answer)
- Instrumentation profiler
- Bytecode rewriting profiler
- Full stack trace profiler
Correct answer: Sampling profiler
Sampling profilers interrupt execution at intervals to record the call stack, adding minimal overhead compared to instrumenting every method call.
Question 6: When tuning a database for read-heavy workloads, which adjustment typically yields the greatest performance gain?
- Increasing the buffer pool or cache size (Correct answer)
- Reducing the number of indexes
- Switching from SSD to HDD storage
- Disabling query result caching
Correct answer: Increasing the buffer pool or cache size
Enlarging the buffer pool keeps frequently accessed data in memory, dramatically reducing disk I/O for read-heavy workloads.
Question 7: A benchmark reports 99th-percentile latency of 2,000 ms while median latency is 50 ms. What does this gap most likely indicate?
- Occasional long garbage collection pauses or lock contention (Correct answer)
- The benchmark tool is misconfigured
- Average throughput is too low
- Network bandwidth is saturated
Correct answer: Occasional long garbage collection pauses or lock contention
A large gap between median and tail latency typically points to intermittent events like GC pauses, lock contention, or resource exhaustion affecting a small fraction of requests.
CSI Performance Benchmarking & Tuning 2
Which metric best indicates whether a system is CPU-bound versus I/O-bound during a benchmark run?