Reading Profiler Output Exercises

1 / 5

A senior engineer opens a CPU flame graph and points to a wide, flat bar at the top: "The flame graph shows a wide bar for the serialize() function."

What does a wide bar in a CPU flame graph indicate?

In a flame graph (invented by Brendan Gregg): the X-axis represents time (proportion of samples — wider = more CPU time); the Y-axis represents call stack depth (higher = deeper). A wide flat bar at the top means a leaf function is consuming that CPU time directly.

Flame graph feature	Meaning
Wide bar (high % of samples)	This function or its children consume significant CPU time — a hot path candidate for optimisation
Tall stack at same width	Deep call chain — all calls are in service of the parent wide bar
Narrow bar	Function uses little CPU — low priority for optimisation
Leaf function	A function at the top of the stack with no children — it is doing the actual CPU work

2 / 5

A profiler reports for the same function: "CPU time: 120ms, Wall time: 340ms."

What does the 220ms difference between CPU time and wall time indicate?

CPU time = actual processor cycles used. Wall time (clock time) = elapsed real time. The difference represents time the thread was blocked: waiting for I/O to complete, waiting to acquire a lock, sleeping, or waiting for a network response.

Pattern	Diagnosis
Wall time ≫ CPU time	I/O-bound — most time is waiting, not computing
CPU time ≈ Wall time	CPU-bound — the thread is actively computing the entire time
Intermittent spikes in wall time	Lock contention or GC pauses causing stop-the-world events

Key vocabulary: I/O bound, CPU bound, lock contention, context switch overhead.

3 / 5

A JVM profiler shows: "GC pause: 1,200ms total — 47 major GC cycles in 10 minutes."

What does this output indicate about the application's performance?

Major GC (stop-the-world) pauses all application threads. 47 pauses in 10 minutes = significant interference with request handling. This contributes directly to high p99 latency and reduced throughput during GC events.

Term	Meaning
stop-the-world (STW)	All application threads paused while GC runs — direct latency hit
GC tuning	Adjusting heap size, GC algorithm, and generation sizing to reduce pause frequency and duration
G1GC / ZGC	Low-latency GC algorithms that reduce STW pause times for latency-sensitive workloads
heap pressure	GC running too frequently because insufficient free heap — investigate memory leak or heap sizing

4 / 5

A profiler shows 60% of total request time in com.example.DAO.findAll() which queries the database.

How should this finding be communicated in a performance review to be most effective?

Effective performance finding communication requires: specific method name, percentage of total time (not just CPU %), root cause identified (not just "slow"), proposed fix, and expected outcome. Vague findings ("the database is slow") don't enable action.

Component	Example from the correct answer
Specific location	`findAll()` method
Measured impact	60% of total request time
Root cause	Unbounded query — fetches all records
Proposed fix + expected outcome	Add pagination + index → <5% of request time

Key vocabulary: hot path, profiling-informed optimisation, estimated impact, unbounded query.

5 / 5

An APM tool flags: "N+1 query problem detected — 2,847 SQL queries executed for a list of 200 orders."

What is the N+1 query problem and how is it identified in profiling output?

The N+1 problem: 1 query fetches 200 orders, then 200 individual queries each fetch the customer for that order = 201 queries total instead of 1 JOIN. In profiling, it appears as: very high query count, short per-query duration, large cumulative overhead.

Term	Meaning
eager loading	Fetch related data in the initial query (JOIN or batch load) — prevents N+1
lazy loading	Fetch related data on demand per object — triggers N+1 in loops
batch preload	Fetch all related records in a single IN query instead of one per object
N+1 detection via APM	APM tools detect repeated similar queries with high count and flag as N+1