STORAGE LATENCY x. RAMAC 350 (600 ms) NAND SSD (60 us)

Transcription

1 1

2 STORAGE LATENCY 2 RAMAC 350 (600 ms) x NAND SSD (60 us) 2016

3 COMPUTE LATENCY 3 RAMAC 305 (100 Hz) x 1000x CORE I7 (1 GHZ) 2016

4 NON-VOLATILE MEMORY 1000x faster than NAND 3D XPOINT (60 ns) x denser than DRAM 1000x endurance of NAND 4

5 NON-VOLATILE MEMORY Support in Linux D XPOINT (60 ns) 2017 New assembly instructions SNIA programming model 5

6 WHAT NVM MEANS FOR DATABASES? 6 Option 1: Treat NVM like a faster SSD DBMS DRAM Use NVM as a logging and backing store Improves performance NVM

7 WHAT NVM MEANS FOR DATABASES? 7 Option 2: Treat NVM as extended memory DBMS DRAM NVM Redesign the key algorithms in a database system Improves not only performance, but also availability and device utilization

8 8 PAST: EXISTING SYSTEMS PRESENT: NVM-AWARE DBMS FUTURE: ANALYTICS ON NVM

9 PAST EXISTING SYSTEMS Investigate how existing systems perform on NVM Option 1: Treat NVM like a faster SSD Evaluate two types of DBMSs Disk-oriented DBMS In-memory DBMS A PROLEGOMENON ON OLTP DATABASE SYSTEMS FOR NON-VOLATILE MEMORY ADMS

10 DBMS ARCHITECTURES 10 DISK-ORIENTED DBMS IN-MEMORY DBMS DRAM Buffer Pool DRAM Table Heap NVM Table Heap NVM Log Checkpoints Log Checkpoints

11 NVM HARDWARE EMULATOR 11 Tunable DRAM latency for emulating NVM Special assembly instructions for NVM Cache line write-back STORE L1 Cache L2 Cache CLWB CPU CACHE NVM

12 EVALUATION 12 Compare existing DBMSs on NVM emulator MySQL (Disk-oriented DBMS) H-Store (In-memory DBMS) TPC-C benchmark 1/8 th of database fits in DRAM, rest on NVM

13 PERFORMANCE 13 Disk-Oriented DBMS In-memory DBMS 40,000 8x DRAM Latency 12x Throughput (txn/sec) 20,000 0 Database systems

14 PERFORMANCE 14 Disk-Oriented DBMS In-memory DBMS 40,000 2x DRAM Latency 1x 4x Legacy database systems are not prepared for NVM Throughput (txn/sec) 20, x Database systems

15 15 PAST: EXISTING SYSTEMS PRESENT: NVM-AWARE DBMS FUTURE: ANALYTICS ON NVM

16 PRESENT NVM-AWARE DBMS Understand changes required in DBMSs to leverage NVM Option 2: Treat NVM as extended memory Rethink key algorithms in database systems Logging and recovery protocol Designed for real-time analytics Multi-versioned database LET S TALK ABOUT STORAGE AND RECOVERY METHODS FOR NON-VOLATILE MEMORY DATABASE SYSTEMS SIGMOD

17 MULTI-VERSIONED DBMS 17 TUPLE ID BEGIN TIMESTAMP END TIMESTAMP PREVIOUS VERSION TUPLE DATA 1 10 X Y Y

18 THOUGHT EXPERIMENT 18 NVM-only storage hierarchy No volatile DRAM DBMS NVM

19 TRADITIONAL STORAGE ENGINE 19 NVM 1 Table Heap Data Data 2 3 Checkpoints Log

20 NON-VOLATILE POINTER 20 POINTER DATA POINTER DATA DRAM DRAM NVM NVM VOLATILE POINTER CRASH: DISAPPEARS NON-VOLATILE POINTER CRASH: VALID ü

21 NVM-AWARE STORAGE ENGINE 21 Instead of duplicating tuple data in log Store non-volatile tuple pointers in log records TRADITIONAL LOG INSERT TUPLE 1 (DATA) UPDATE TUPLE 1 (DATA) NVM-AWARE LOG INSERT TUPLE 1 (POINTER) UPDATE TUPLE 1 (POINTER)

22 NVM-AWARE STORAGE ENGINE 22 NVM Meta Data 2 Log 1 Table Heap Checkpoints

23 EVALUATION 23 Compare storage engines on NVM emulator Traditional engine NVM-aware engine Yahoo! Cloud Serving Benchmark Database fits in NVM

24 PERFORMANCE 24 Traditional Engine NVM-Aware Engine 600,000 8x DRAM Latency 3x Throughput (txn/sec) 300,000 0 Storage Engines

25 PERFORMANCE 25 Throughput (txn/sec) Traditional Engine NVM 1,500,000 latency has a significant impact on the 750,000 2x DRAM Latency 6x NVM-Aware Engine 4x performance of NVM-aware storage engine 1x 0 Storage Engines

26 DEVICE LIFETIME 26 Traditional Engine NVM-Aware Engine NVM Stores (M) x 0 Storage Engines

27 THE PELOTON DBMS A new database system for NVM research Designed for a two-tier storage hierarchy DBMS DRAM WRITE-BEHIND LOGGING UNDER REVIEW NVM 27

28 WRITE-BEHIND LOGGING 28 Write-ahead log serves two purposes Transform random database writes into sequential log writes Support transaction rollback NVM supports fast random writes Directly write data to the database Later, record metadata in write-behind log

29 WRITE-BEHIND LOGGING 29 DRAM Meta Data 1 Database Data NVM 3 Log 2 Database

30 METADATA FOR INSTANT RECOVERY 30 Record failed timestamp ranges in log Use it to ignore effects of uncommitted transactions Group Commit T 1 T 2 T 3 T 4 (T 1, T 2 ) (T 1, T 2 ) Time Dirty Ranges Garbage Collection

31 EVALUATION 31 Compare logging protocols in Peloton Write-Ahead logging Write-Behind logging Storage devices in Intel s NVM hardware emulator Hard-disk drive Solid-state drive Non-volatile memory emulator TPC-C benchmark

32 APPLICATION AVAILABILITY 32 Write-Ahead Logging Write-Behind Logging 10,000 Recovery Time (sec) Hard Disk Drive Solid State Drive Non-Volatile Memory 1000x

33 PERFORMANCE 33 Write-Ahead Logging Write-Behind Logging 10, x Throughput (txn/sec) x 1 Hard Disk Drive Solid State Drive Non-Volatile Memory

34 34 PAST: EXISTING SYSTEMS PRESENT: NVM-AWARE DBMS FUTURE: ANALYTICS ON NVM

35 FUTURE REAL-TIME ANALYTICS Support analytics on larger-than-nvm databases Cost of first-generation NVM devices Three-tier storage hierarchy DRAM + NVM + SSD When to migrate data between different layers? REAL-TIME ANALYTICS IN NON-VOLATILE MEMORY DATABASE SYSTEMS WORK IN PROGRESS 35

36 THREE-TIER STORAGE HIERARCHY 36 DRAM Meta Data 1 Database Data NVM 3 Log 2 Database Data SSD 4 Database

37 DATA MIGRATION 37 Can directly read warm data from NVM No need to copy data over to DRAM for reading Keep hottest data in DRAM Dynamically migrate cold data to SSD

38 THREE-TIER STORAGE HIERARCHY 38 Throughput (txn/sec) 10% on NVM 50% on NVM 90% on NVM 120,000 60,000 0 Read-Heavy Workload 3x Balanced Workload 4x Write-Heavy Workload 8x

39 39 PAST: EXISTING SYSTEMS PRESENT: NVM-ORIENTED NVM-AWARE DBMS FUTURE: ANALYTICS ON NVM

40 LESSONS LEARNED 40 NVM outperforms SSD when correctly used Rethink key algorithms in database systems Write-behind logging enables instant recovery Improves device utilization and extends its lifetime Ongoing work Data placement policy Replication

41 41

42 42