1. PostgreSQL 성능을 위해 개발된
라이브러리 OS 소개

2. Special application on top of general OS (Linux)
Problem
2
Linux
PostgreSQL
Hardware
General purpose: Web server, office, game, …
Bottleneck for DB performance
Special need: ACID, high concurrency, fast data access…

3. AppOS
Solution
Linux
PostgreSQL
Hardware
Specialized for DB performance
Special need: ACID, high concurrency, fast data access…
Special application on top of special OS (AppOS)
3

4. Solution
V1
Special application on top of special OS (AppOS)
Modified
Linux
PostgreSQL
Hardware
V2
Linux
PostgreSQL
Hardware
Module
Module
V3
Linux
PostgreSQL
Hardware
AppOS
4

5. Merits
5
Portable Performant Extensible

6. AppOS provides
system call APIs
PostgreSQL is built
on top of Linux system calls
libhook
libapp
libcore
syscall
PostgreSQL
AppOS
AppOS seamlessly runs
by hooking system calls
open() read()...
libhook
libapp
libcore
PostgreSQL
AppOS
open() read()...
6
AppOS doesn’t require modifications
Technology: Portability

7. Technology: Portability
AppOS is highly portable because it is built on top of Linux ABIs
LinuxLinux Linux
LinuxLinux
7
AppOS runs in diverse environments

8. Technology: Performance
Virtual File System
libcore
libapp
libhook
AppOS
Page Cache
Physical File Mapping
mongodbpostgresql
Memory
Mgmt
Utility
…
I/O Engine
• libapp
• Per-DB optimizations
• libcore
• DB-common optimizations
8
AppOS consists of user-level libraries

9. Technology: Performance
9
Linux doesn’t know DB-internal semantics

10. Technology: Performance
AppOS infers semantics by observing system call patterns
10
AppOS understands DB-internal semantics

11. Technology: Performance
write() write()
T2
Linux File System
MetadataFile Journal
Linux Page Cache
T1
Page
T3
T4
rename()
T5
write()
write()
Only one thread
can write to a file
Complex locking
behaviors for file
system consistency
11
Linux imposes unnecessary overheads for file accesses

12. Technology: Performance
write() write()
T2
Linux File System
AppOS Page Cache
T1
Page
T4
rename()
T5
write()
Multiple threads
can concurrently
write to different
positions of a file
Simple locking based
on data consistency
guarantees provided
by DB
Page
Asynchronous
direct I/Os by
I/O workers
12
AppOS decouples page cache from file system

13. Technology: Performance
13
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer
Abstraction

14. Technology: Performance
14
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer
Abstraction
Buffer Cache
read() write()
admission
control

15. Technology: Performance
15
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer
Abstraction
Buffer Cache
read() write()
admission
control
Block-level Q

16. Technology: Performance
16
Linux I/O path delays high-priority disk I/Os
Application
Storage Device
Caching Layer
File System Layer
Block Layer
Abstraction
Buffer Cache
reorder
FG FG BGBG
read() write()

17. Technology: Performance
17
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer
Abstraction
Buffer Cache
read() write()
FG FGBG
Device-internal Q
admission
control

18. Technology: Performance
18
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer
Abstraction
Buffer Cache
read() write()
FG FGBG
BG FG BGBG
reorder

19. Technology: Performance
19
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer
Locks
Condition variables

20. Technology: Performance
20
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer Condition variables
I/OFG
lock
BG
wait

21. Technology: Performance
21
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer
I/OFG
lock
BG
wait
FG
wait
wait
BGvar
wake

22. Technology: Performance
22
Linux I/O path delays high-priority disk I/Os
Storage Device
Caching Layer
Application
File System Layer
Block Layer
I/O
FG
wait
wait
BGuser
var
wake
FG
wait

23. Technology: Performance
AppOS libapp
Storage
(I/O class, share)(latency, 1000) (throughput, 500)(latency, 1000)
AppOS I/O Engine
WAL write
Foreground
data read
Background read/write
(e.g., checkpoint, compaction)
Dispatch I/Os based on priority only if
there is no storage congestion
23
AppOS schedules I/Os based on DB-internal priority

24. Technology: Performance
Linux
Cache
PostgreSQL
Cache
Block 1
Data block
Block 2
Storage
1. Crash
happens
Block 1 Block 2
Torn page problem
Block 1
Data block
Block 2
2. Crash happens
Block 1 Block 2 WAL
Full page write
1. Write block
to WAL file
3. Recover upon reboot
2. Data block is corrupted
24
Linux doesn’t support atomic write
2X
Write

25. Technology: Performance
AppOS manages file mapping for eliminating write amplification
25
AppOS supports atomic write

26. Technology: Performance
26
Benchmarks show better performance for OLTP
* Host
- AWS ec2
- c4.8xlarge
- 36 cores, 60GB mem
- 1000GB SSD-backed EBS
- PostgreSQL 10.3
- 24GB shared buffers
- 10GB max wal
* Client
- AWS ec2
- c4.xlarge
- 4 cores, 7.5GB mem
- OLTP-Bench
- TPC-C with 200 scale factor
(전자상거래 트랜잭션 처리)

27. Technology: Performance
27
Benchmarks show better performance for OLTP
* Host
- AWS ec2
- c4.8xlarge
- 36 cores, 60GB mem
- 1000GB SSD-backed EBS
- PostgreSQL 10.3
- 24GB shared buffers
- 10GB max wal
* Client
- AWS ec2
- c4.xlarge
- 4 cores, 7.5GB mem
- OLTP-Bench
- TPC-C with 200 scale factor
(전자상거래 트랜잭션 처리)

28. Technology: Extensibility
PostgreSQLMongoDB
: 70 syscalls : 74 syscalls
Common
: 50 syscalls
28
Used system calls are similar from DB to DB

29. Technology: Extensibility
PostgreSQL
9.5
PostgreSQL
9.6
PostgreSQL
10.0
poll() -> epoll()
semop() -> futex()
29
Used system calls are not changed a lot

30. Technology: Extensibility
libcore
libapp
libhook
AppOS
postgresql
PostgreSQL
libcore
libapp
libhook
AppOS
mongodb
MongoDB
libcore
libapp
libhook
AppOS
mysql
MySQL
libcore
libapp
libhook
AppOS
redis
Redis
...
30
AppOS can easily support various databases

31. Roadmap
31
Efficient cache management
Linux
Cache
PostgreSQL
Cache
Block 1 Block 2
Storage
read
Block 1 Block N
Duplicated LRU
Block 1 Block 2
Block 1 Block N
DB cache-aware
Block 1
next victim
read
Block 3 Block 1 Block 3
next victimread
read

32. Roadmap
32
Parallel query optimization

33. Roadmap
33
Parallel query optimization

34. Roadmap
34
Parallel query optimization

35. Roadmap
35
Cooperative CPU scheduling

36. Vision
36
Support diverse types of applications

37. https://apposha.io
sangwook@apposha.io