Kako optimizirate zmogljivost JVM aplikacij?

Question

Accepted Answer

Optimizacija JVM aplikacij obsega **profiliranje** za iskanje ozkih grl, **tuning JVM/GC**, **izboljšave na ravni kode** in izkoriščanje JIT prevajalnika. Kot vedno velja pravilo: **najprej izmeri, nato optimiziraj dokazano ozko grlo** — nikoli ne ugibaj.

## Profiliraj pred optimizacijo

```text
Tools:
  ✓ JProfiler / YourKit / VisualVM — CPU & memory profiling, hotspots
  ✓ Java Flight Recorder (JFR) + Mission Control — low-overhead production profiling
  ✓ async-profiler — flame graphs for CPU/allocation
  ✓ heap dumps + Eclipse MAT — memory leak analysis
```

Profiliranje razkrije *kje* se čas in pomnilnik dejansko potrošita, zato popraviš pravo 20 % ki je važna, namesto da mikro-optimiziraš nepomembno kodo.

## JIT prevajalnik dela zate

```text
The JVM's JIT (Just-In-Time) compiler identifies "hot" frequently-run code and
compiles it to optimized native machine code at runtime (with inlining, etc.).
→ Java often approaches native speed for hot paths after warm-up.
→ Implication: benchmark AFTER warm-up; let the JIT optimize before measuring.
```

## GC tuning

```bash
# choose a collector appropriate to the workload, size the heap
java -Xms2g -Xmx2g \          # set heap size (avoid resizing; -Xms = -Xmx)
     -XX:+UseG1GC \           # G1 (balanced, default) — or ZGC for low pause
     -XX:MaxGCPauseMillis=200 \
MyApp
```

```text
✓ Right-size the heap (too small → frequent GC; too large → long pauses)
✓ Pick the GC for your goal: G1 (balanced), Parallel (throughput), ZGC (low latency)
✓ Reduce allocation churn — fewer short-lived objects = less GC pressure
```

## Optimizacije na ravni kode

```text
✓ Better algorithms & data structures — biggest wins (HashMap O(1) vs list O(n))
✓ Avoid unnecessary object creation (esp. in loops; reuse, use primitives over wrappers)
✓ Use StringBuilder for string building in loops (not + concatenation)
✓ Minimize autoboxing (Integer vs int) in hot paths
✓ Cache expensive results; use connection/thread pools
✓ Lazy initialization for costly objects not always needed
✓ Reduce I/O round-trips; batch database queries (avoid N+1)
```

## Baza podatkov je pogosto pravo ozko grlo

```text
For server apps, the slow part is frequently DB queries (N+1, missing indexes) or
network I/O — NOT Java code. Profiling reveals this so you fix the true cause.
```

## Izogni se prezgodnji optimizaciji

```text
"Premature optimization is the root of all evil." Write clear, correct code first;
profile under realistic load; then optimize the measured hotspots specifically.
```

## Zakaj je to pomembno

Optimizacija zmogljivosti JVM je visoko vredna, vendar mora biti **disciplinirana in vodena z meritvami** — ugibanje zapravi trud in dodaja kompleksnost.

Znanje celotnega nabora orodij je pomembno: profilerji (JFR, async-profiler) za lociranje resničnih ozkih grl; razumevanje JIT (ki optimizira vroč kod med izvajanjem, zato benchmarki potrebujejo ogrevanje); GC tuning (velikost heap in izbira kolektorja — G1/ZGC — za uravnoteženost med prepustnostjo in časom zaustavitve); in zmage na ravni kode (boljši algoritmi/podatkovne strukture, zmanjšanje alokacij in samodejnega pakiranja, pooling, batching).

Ugotovitev, da je ozko grlo pogosto **baza podatkov ali I/O**, ne Java koda, pravilno usmerja trud.

Ta kombinacija — profiliraj najprej, tuniraj JVM in GC primerno, optimiziraj dokazana vroča mesta in izogni se prezgodni mikro-optimizaciji — je tisto, kar drži velike JVM aplikacije hitre in razširljive, in odraža strokovno znanje na višji ravni.