반응형
개념 차이
AIO (Linux Native AIO)
- Linux의 전통적인 비동기 I/O 인터페이스 (libaio)
- 커널에 I/O 요청을 제출하고 완료를 polling 또는 event 방식으로 수신
- 구조적으로는 비동기지만 실제 내부는 제한적 비동기
👉 핵심 특징
- syscall 기반 (io_submit, io_getevents)
- 제한적인 operation 지원 (주로 direct I/O)
- completion polling 비용 존재
io_uring
- Linux 5.1+에서 도입된 차세대 I/O 인터페이스
- Submission Queue (SQ) / Completion Queue (CQ) 기반 공유 메모리 구조
- syscall을 최소화하고 zero-copy + batch 처리 최적화
👉 핵심 특징
- ring buffer 기반 (user ↔ kernel 공유)
- 거의 모든 I/O 지원 (file, network, fsync 등)
- polling 없이도 high performance 가능
KVM / QEMU 관점에서 비교
KVM 자체는 하이퍼바이저이고, 실제 I/O는 QEMU + host kernel에서 처리됩니다.
즉 비교 포인트는 “QEMU block layer에서 어떤 I/O 엔진을 쓰느냐”입니다.
구조 비교
AIO (기존 방식)
Guest VM
→ virtio-blk/scsi
→ QEMU
→ libaio
→ kernel block layer
→ virtio-blk/scsi
→ QEMU
→ libaio
→ kernel block layer
io_uring
Guest VM
→ virtio-blk/scsi
→ QEMU
→ io_uring
→ kernel (shared ring)
→ virtio-blk/scsi
→ QEMU
→ io_uring
→ kernel (shared ring)
성능 비교
(1) Latency
- AIO: syscall overhead + context switch 존재
- io_uring: syscall 최소화 → latency 감소
👉 결과
- io_uring이 10~30% 낮은 latency (워크로드에 따라 더 큼)
(2) Throughput (IOPS)
- AIO: 큐 depth 증가 시 병목 발생
- io_uring: batch + lockless 구조
👉 결과
- io_uring이 고QD 환경에서 압도적 (특히 NVMe)
(3) CPU 사용량
- AIO: syscall + wakeup 비용 큼
- io_uring: polling + shared memory
👉 결과
- io_urin
(4) Scaling (멀티 VM 환경)
- AIO: contention 발생
- io_uring: per-ring 구조 → 확장성 우수
io_uring 기반 vm 생성
sudo /usr/bin/qemu-system-aarch64 -enable-kvm -cpu host -smp 4 -m 4096 -machine virt -bios /usr/share/qemu-efi-aarch64/QEMU_EFI.fd -drive file=/home/sunshout/vm/jammy-server-cloudimg-arm64.img,format=qcow2,if=virtio,aio=io_uring -nographic -serial mon:stdio -drive file=seed.iso,format=raw,if=virtio -netdev tap,id=net0,ifname=vnet0,script=no,downscript=no -device virtio-net-pci,netdev=net0,mac=52:54:00:12:34:57aio 기반 vm 생성
sysbench 성능 측정
# 1. 2GB 테스트 파일 생성 (이제 No space left 에러 없이 진행될 겁니다)
sysbench fileio --file-total-size=2G prepare
# 2. 성능 측정 실행 (4개의 스레드, 비동기 모드)
sysbench fileio --file-total-size=2G \
--file-test-mode=rndrw \
--num-threads=4 \
--file-io-mode=async \
--file-extra-flags=direct \
--time=60 \
run
io_uring
Guest multiqueue + Host single queue (io_uring)
File operations:
reads/s: 23107.98
writes/s: 15406.87
fsyncs/s: 22806.78
Throughput:
read, MiB/s: 361.06
written, MiB/s: 240.73
General statistics:
total time: 60.0107s
total number of events: 3679596
Latency (ms):
min: 0.00
avg: 0.06
max: 13.60
95th percentile: 0.08
sum: 235822.23
Threads fairness:
events (avg/stddev): 919899.0000/2698.18
execution time (avg/stddev): 58.9556/0.00Guest multiqueue + Host multi queue (iothread)
File operations:
reads/s: 30538.67
writes/s: 20362.64
fsyncs/s: 30300.34
Throughput:
read, MiB/s: 477.17
written, MiB/s: 318.17
General statistics:
total time: 60.0038s
total number of events: 4872100
Latency (ms):
min: 0.00
avg: 0.05
max: 24.47
95th percentile: 0.06
sum: 235004.18
Threads fairness:
events (avg/stddev): 1218025.0000/5765.36
execution time (avg/stddev): 58.7510/0.01
File operations:
reads/s: 29887.46
writes/s: 19925.78
fsyncs/s: 29716.34
Throughput:
read, MiB/s: 466.99
written, MiB/s: 311.34
General statistics:
total time: 60.0034s
total number of events: 4771751
Latency (ms):
min: 0.00
avg: 0.05
max: 38.71
95th percentile: 0.06
sum: 234999.41
Threads fairness:
events (avg/stddev): 1192937.7500/3808.45
execution time (avg/stddev): 58.7499/0.00
aio
Guest multiquque + Host single queue
File operations:
reads/s: 21936.94
writes/s: 14625.35
fsyncs/s: 21806.89
Throughput:
read, MiB/s: 342.76
written, MiB/s: 228.52
General statistics:
total time: 60.0048s
total number of events: 3502058
Latency (ms):
min: 0.00
avg: 0.07
max: 27.57
95th percentile: 0.07
sum: 234956.16
Threads fairness:
events (avg/stddev): 875514.5000/3233.71
execution time (avg/stddev): 58.7390/0.00Guest multiqueue + Host multi queue
File operations:
reads/s: 28671.55
writes/s: 19114.65
fsyncs/s: 28418.94
Throughput:
read, MiB/s: 447.99
written, MiB/s: 298.67
General statistics:
total time: 60.0037s
total number of events: 4572160
Latency (ms):
min: 0.00
avg: 0.05
max: 28.39
95th percentile: 0.08
sum: 235281.31
Threads fairness:
events (avg/stddev): 1143040.0000/4828.72
execution time (avg/stddev): 58.8203/0.00
File operations:
reads/s: 29058.11
writes/s: 19371.99
fsyncs/s: 28807.90
Throughput:
read, MiB/s: 454.03
written, MiB/s: 302.69
General statistics:
total time: 60.0041s
total number of events: 4634237
Latency (ms):
min: 0.00
avg: 0.05
max: 19.13
95th percentile: 0.07
sum: 235252.50
Threads fairness:
events (avg/stddev): 1158559.2500/1889.27
execution time (avg/stddev): 58.8131/0.00728x90
반응형