Ni nini kinachozuia concurrency kwenye scale? Eleza contention, false sharing, na backpressure.

Question

IT Interview Practice · Accepted Answer

**Kwenye scale, throughput inazuiwa kidogo na idadi ya CPU kuliko na contention, athari za cache, na queues zisizo na kikomo — kuongeza threads baada ya kikomo fulani hufanya mambo kuwa polepole zaidi.** Amdahl's na Universal Scalability Law zote zinatabiri hili: gharama ya uratibu hatimaye inatawala.

## Contention

Wakati threads nyingi zinagombea lock moja au cache line, zinakuwa serial na kutumia muda kusubiri badala ya kufanya kazi. Chini ya Universal Scalability Law, contention pamoja na **coherency cost** inaweza kufanya throughput *kupungua* baada ya idadi bora ya threads.

**Njia za kupunguza**: shard/partition data, state ya kila-thread (thread-local, striped counters kama `LongAdder`), read-write locks, au lock-free structures.

## False sharing

Variables mbili huru ambazo zinatokea kukaa kwenye **cache line ile ile ya CPU** husababisha cores kubatilisha caches za kila mmoja kwa kila write, ingawa kamwe hazigusi data ile ile — kupungua kwa kasi kimyakimya na kwa ukatili.

```text
cache line (64B): [ counterA | counterB ]  ← two cores ping-pong the line
Fix: pad so each hot variable owns its own line
```

```java
@Contended long counterA;   // JVM pads to a separate cache line
```

## Backpressure

Wakati producers zinapita consumers kwa kasi, queue isiyo na kikomo inakua hadi OOM na latency inalipuka. **Backpressure** huashiria producers kupunguza kasi: bounded queues zinazozuia au kukataa, credit-based flow control, au reactive-streams `request(n)`.

```text
producer ──▶ [bounded queue full] ──▶ block/reject ──▶ producer slows
```

## Kwa nini ni muhimu

Hizi ni athari zinazotenganisha "inafanya kazi kwenye cores 4" na "inaongezeka hadi 64." Designs za kijinga zinagonga ukuta: locks zinakuwa serial, cache lines zinakwaruzana, queues zinavimba. Sharding kupunguza contention, padding kuua false sharing, na backpressure kuweka kikomo cha queues ndio vifunguo vinavyoweka latency na throughput vyenye afya chini ya mzigo halisi.

Contention

False sharing

Backpressure

Kwa nini ni muhimu

Maswali yanayohusiana