LLM 블랙박스 해부하기: JProfiler의 사용자 지정 분류
현대의 에이전틱 애플리케이션은 다양한 논리적 작업을 단일 인터페이스를 통해 수행하는 경우가 많습니다. 이러한 작업들은 성능과 비용 특성이 크게 다르지만, 기존 프로파일러에서는 구분 없는 하나의 호출 트리로만 나타납니다.
기본 설정에서 JProfiler는 이러한 요청을 모델 이름을 기준으로 그룹화합니다. 이를 통해 전체적인 개요는 파악할 수 있지만, 서로 다른 기능별 워크로드를 구분하는 데 필요한 세분성은 부족합니다.
이 스크린캐스트에서는 내부 메타데이터, 지시 구조 또는 애플리케이션 상태를 기반으로 사용자 지정 분류 규칙을 정의하는 스크립트를 통해 이러한 사전 정의된 관점에서 한 단계 더 나아가는 방법을 보여줍니다.
AI 프로파일링: LangChain4j와 Spring AI
에이전틱 애플리케이션은 고유한 프로파일링 과제를 안겨줍니다. 표준적인 CPU 사용량을 넘어, 성능과 비용은 프레임워크 추상화 뒤에 숨어 있는 프롬프트, RAG 리트리버, 도구 호출의 복잡한 구조에 의해 결정됩니다.
벤더 종속성을 피하기 위해 대부분의 개발자는 LangChain4j나 Spring AI 같은 프레임워크를 사용합니다. JProfiler의 AI 프로브는 이러한 프레임워크 내부를 깊이 있게 들여다볼 수 있게 해 줍니다.
이 스크린캐스트는 LangChain4j 고객 지원 에이전트의 프로파일링 과정을 안내하면서, AI 프로브가 프롬프트 구성을 시각화하고 토큰을 많이 소모하는 이상값을 분리해 내며 리소스 비용을 기록된 호출 트리에 직접 투영하는 방법을 보여줍니다.
Heap Walker scripting for snapshot analysis
Previously, JProfiler's heap walker only offered advanced filtering and grouping features in live sessions. With JProfiler 15, the heap walker introduces full support for filtering and grouping in snapshots. A new API lets you write scripts to filter objects or define custom grouping logic based on dumped object data.
This screencast demonstrates the new functionality with a small Hibernate-based demo that creates different types of query cache entries. Hibernate’s second-level cache can quickly fill up with unexpected entries, and identifying what’s taking up space is key to resolving memory issues. JProfiler now makes it possible to answer such questions with snapshots.
The screencast shows how you can:
- Group cache keys by query structure to compare memory impact
- Use object set filter scripts to isolate problematic cache entries
- Use merged dominating references to understand why objects are retained
Capturing and comparing MBean states with JProfiler
MBeans expose a wide range of runtime metrics and management operations that are essential for understanding the internal state of Java applications. Traditionally, JProfiler has supported live inspection of MBeans during a profiling session.
With the new MBean snapshot feature in JProfiler 15, it is now possible to capture the complete MBean state, compare snapshots, and analyze metrics offline, even after the application has shut down.
This screencast demonstrates the new functionality with a small Hazelcast application that performs random map operations and registers a custom MBean for triggering cleanup operations.
Along the way, you will see how to:
- Capture and compare MBean snapshots manually
- Trigger MBean snapshots programmatically with method triggers
- Restrict the snapshot scope to specific MBeans
Advanced SSH remote profiling with JProfiler
Remote profiling can be challenging when SSH access involves custom setups, proxies, or cloud authentication plugins. JProfiler 15 introduces advanced SSH connection features that remove these obstacles, enabling profiling in complex environments without manual setup.
This screen cast shows how to profile remote JVMs even in situations where the built-in SSH client is insufficient, using AWS Session Manager as an example.
Along the way, you will see how to:
- Use OpenSSH mode for remote profiling
- Connect with ProxyCommand and the AWS CLI
- Handle SSH port forwarding restrictions with netcat mode
Profiling Java applications in VS Code with JProfiler
JProfiler now integrates directly into VS Code, bringing the powerful profiling tool into your everyday development workflow.
In this screencast, we profile the Spring Pet Clinic demo and walk through the features of the JProfiler extension for VS Code, including:
- Toggling profiling mode
- Automatic setup of profiled packages
- Source code navigation from profiling views
- Attach mode for unsupported launch methods (like Maven)
- Built-in Gradle integration
If you're using a launch method in VS Code that isn't currently supported, let us know, we’re always looking to expand support.
Advanced Kafka probe configuration in JProfiler
In our first Kafka screencast, we looked at how JProfiler correlates message flow and CPU usage through producers, consumers, and topic-level call trees.
In this follow-up, we go deeper, showing how to configure custom naming for Kafka messages based on their content.
This enables you to:
- split call trees by semantic message types
- track specific event types in probe views
- add value-level details to single event descriptions
- visualize outliers in event duration histograms
We also trigger a Kafka message spike via an API call and analyze its impact, all without changing the application code.
Profiling Kafka #1 – Message Flow & Hot Spots
Kafka-based applications are notoriously hard to profile. Message flow happens in background threads, stack traces don’t reveal the full picture, and it’s hard to know which topic contributes to which hot spot.
JProfiler’s Kafka probes give you deep insight: producers, consumers, and message-level details are all captured and correlated with performance data.
This screencast walks through profiling a Spring Kafka application, showing how:
- consumer call trees are split by topic
- producers link directly into probe views
- you can correlate CPU hot spots with Kafka topics
- Kafka events are filterable, inspectable, and aggregatable
Garbage collector analysis in JProfiler
This screencast shows how to use the garbage collector probe in JProfiler. Having access to detailed information about the overall activity of the GC, as well as the single garbage collections, is crucial for tuning the garbage collector and achieving an optimal performance for your application.
Recording JFR snapshots with JProfiler
Recording JFR snapshots with JProfiler
This screencast shows JProfiler's versatile functionality as a JFR recording controller. As an example, a JFR recording on a Kubernetes cluster is recorded and the resulting snapshot is shown in JProfiler. In this context, you can see the wizard for configuring JFR recording settings. In addition, JFR recordings of terminated JVMs and the handling of externally started JFR recordings are demonstrated.