DHTM: Durable Hardware Transactional Memory Arpit Joshi , Vijay - PowerPoint PPT Presentation

DHTM: Durable Hardware Transactional Memory Arpit Joshi , Vijay Nagarajan, Marcelo Cintra, Stratis Viglas ISCA 2018

Persistent Memory is here… � 2

Persistent Memory Systems L1 L1 LLC Persistent Memory � 3

Persistent Memory Systems L1 L1 • Persistent Memory - Non-volatility over the memory bus - Load/Store interface to persistent data LLC Persistent Memory � 3

Persistent Memory Systems L1 L1 • Persistent Memory - Non-volatility over the memory bus - Load/Store interface to persistent data LLC System Crashes Persistent Memory � 3

Persistent Memory Systems L1 L1 • Persistent Memory - Non-volatility over the memory bus - Load/Store interface to persistent data LLC System Crashes • Crash Consistency - Is the persistent state consistent? Persistent Memory - Programming Model: ACID Transactions � 3

Persistent Memory Systems L1 L1 • Persistent Memory “Ensuring failure atomicity for all this computation without - Non-volatility over the memory bus failure-atomic transactions is practically infeasible, if not - Load/Store interface to persistent data impossible.” LLC System Crashes Marathe et al. [HotStorage’17] • Crash Consistency - Is the persistent state consistent? Persistent Memory - Programming Model: ACID Transactions � 3

Persistent Memory Systems L1 L1 • Persistent Memory “Ensuring failure atomicity for all this computation without - Non-volatility over the memory bus failure-atomic transactions is practically infeasible, if not - Load/Store interface to persistent data impossible.” LLC System Crashes Marathe et al. [HotStorage’17] • Crash Consistency - Is the persistent state consistent? Persistent Memory - Programming Model: ACID Transactions How fast can we support ACID? � 3

ACID Transactions L1 L1 LLC Persistent Memory � 4

ACID Transactions Atomic Visibility L1 L1 LLC Persistent Memory � 4

ACID Transactions Atomic Visibility L1 L1 LLC Persistent Memory Atomic Durability � 4

ACID Transactions Atomic Visibility L1 L1 STM Locks HTM LLC Persistent Memory Atomic Durability � 4

ACID Transactions Atomic Visibility L1 L1 STM Locks HTM LLC Check- H/W S/W Logging pointing Logging Persistent Memory Atomic Durability � 4

Atomic Visibility: HTM � 5

Atomic Visibility: HTM • Commercial HTMs [Intel, IBM] L1 Cache Cache Line R W A = 15 1 B = 20 1 � 5

Atomic Visibility: HTM • Commercial HTMs [Intel, IBM] L1 Cache - Version Management : read/write sets in L1 cache Cache Line R W A = 15 1 B = 20 1 � 5

Atomic Visibility: HTM • Commercial HTMs [Intel, IBM] L1 Cache - Version Management : read/write sets in L1 cache Cache Line R W - Conflict Detection : piggy back on the A = 15 1 coherence protocol B = 20 1 � 5

Atomic Visibility: HTM • Commercial HTMs [Intel, IBM] L1 Cache - Version Management : read/write sets in L1 cache Cache Line R W - Conflict Detection : piggy back on the A = 15 coherence protocol B = 20 - Commit : make updates non-speculative � 5

Atomic Visibility: HTM • Commercial HTMs [Intel, IBM] L1 Cache - Version Management : read/write sets in L1 cache Cache Line R W - Conflict Detection : piggy back on the coherence protocol B = 20 - Commit : make updates non-speculative - Abort : invalidate write set � 5

Atomic Visibility: HTM • Commercial HTMs [Intel, IBM] L1 Cache - Version Management : read/write sets in L1 cache Cache Line R W - Conflict Detection : piggy back on the coherence protocol B = 20 - Commit : make updates non-speculative - Abort : invalidate write set Write-sets in commercial HTMs limited by the size of the L1 cache. � 5

Atomic Durability: Logging � 6

Atomic Durability: Logging Persistent Memory • Logging for durability [ Doshi’16, Joshi’17, Shin’17, Ogleari’18 ] In-place Values A = 10 B = 20 C = 30 � 6

Atomic Durability: Logging Persistent Memory • Logging for durability [ Doshi’16, Joshi’17, Shin’17, Ogleari’18 ] In-place Values Transaction Log A = 10 A = 15 - Write a log entry for every update B = 20 B = 25 C = 30 � 6

Atomic Durability: Logging Persistent Memory • Logging for durability [ Doshi’16, Joshi’17, Shin’17, Ogleari’18 ] In-place Values Transaction Log A = 15 - Write a log entry for every update B = 25 - Commit : Update the values in-place C = 30 � 6

Atomic Durability: Logging Persistent Memory • Logging for durability [ Doshi’16, Joshi’17, Shin’17, Ogleari’18 ] In-place Values Transaction Log A = 15 A = 10 - Write a log entry for every update B = 25 B = 20 - Commit : Update the values in-place C = 30 - Abort : Undo any in-place updates � 6

Atomic Durability: Logging Persistent Memory • Logging for durability [ Doshi’16, Joshi’17, Shin’17, Ogleari’18 ] In-place Values Transaction Log A = 15 A = 10 - Write a log entry for every update B = 25 B = 20 - Commit : Update the values in-place C = 30 - Abort : Undo any in-place updates In-place updates in the critical path of commit High memory write bandwidth requirement � 6

ACID = HTM + Logging Goals: - Support fast commits - Minimise memory bandwidth consumption - Extend the supported transaction size - Maintain the simplicity of commercial HTMs � 7

DHTM: Durable Hardware Transactional Memory L1 L1 Log Writes LLC Persistent Memory � 8

DHTM: Durable Hardware Transactional Memory L1 L1 Commercial HTM + Hardware Redo Log Log Writes LLC Persistent Memory � 8

DHTM: Durable Hardware Transactional Memory L1 L1 Commercial HTM + Hardware Redo Log - H/W Redo Log + Log Buffer Log Writes Reduced memory bandwidth LLC Fast commits Persistent Memory � 8

DHTM: Durable Hardware Transactional Memory L1 L1 Commercial HTM + Hardware Redo Log - H/W Redo Log + Log Buffer Log Writes Reduced memory bandwidth LLC Fast commits - H/W Log + Sticky State Extended transaction size to the LLC Simplicity of commercial HTM Persistent Memory � 8

DHTM: Log Buffer L1 L1 Log Writes LLC Persistent Memory � 9

DHTM: Log Buffer L1 L1 • Redo Log Bandwidth Problem Log Writes LLC Persistent Memory � 9

DHTM: Log Buffer L1 L1 • Redo Log Bandwidth Problem - write a log entry for every store Log Writes LLC Persistent Memory � 9

DHTM: Log Buffer L1 L1 • Redo Log Bandwidth Problem - write a log entry for every store Log Writes - multiple stores create multiple log entries LLC Persistent Memory � 9

DHTM: Log Buffer L1 L1 • Redo Log Bandwidth Problem - write a log entry for every store Log Writes - multiple stores create multiple log entries LLC • Solution: Log Buffer Persistent Memory � 9

DHTM: Log Buffer L1 L1 • Redo Log Bandwidth Problem - write a log entry for every store Log Writes - multiple stores create multiple log entries LLC • Solution: Log Buffer - track cache lines being modified Persistent Memory � 9

DHTM: Log Buffer L1 L1 • Redo Log Bandwidth Problem - write a log entry for every store Log Writes - multiple stores create multiple log entries LLC • Solution: Log Buffer - track cache lines being modified - multiple writes coalesced in a log entry Persistent Memory � 9

DHTM: Log Buffer L1 L1 • Redo Log Bandwidth Problem - write a log entry for every store Log Writes - multiple stores create multiple log entries LLC • Solution: Log Buffer - track cache lines being modified - multiple writes coalesced in a log entry - log entry written to persistent memory on eviction Persistent Memory from log buffer � 9

DHTM: Transaction States � 10

DHTM: Transaction States Begin Transaction Active � 10

DHTM: Transaction States End Transaction & Begin Log Records Transaction Persisted Active Commit � 10

DHTM: Transaction States End Transaction & Begin In-place Data Log Records Transaction Persisted Persisted Commit Active Commit Complete � 10

DHTM: Transaction States End Transaction & Begin In-place Data Log Records Transaction Persisted Persisted Commit Active Commit Complete Abort Conflict � 10

DHTM: Commit Example State L1 Cache Cache Line R W Begin_Transaction Log Buffer Write (A=15) Read (B) Persistent Memory Write (B=25) Transaction Log End_Transaction In-place Values A = 10 A = 10 A = 15 B = 20 B = 25 B = 20 C = 30 � 11

DHTM: Commit Example State L1 Cache Active Cache Line R W Begin_Transaction Log Buffer Write (A=15) Read (B) Persistent Memory Write (B=25) Transaction Log End_Transaction In-place Values A = 10 A = 10 A = 15 B = 20 B = 25 B = 20 C = 30 � 11

DHTM: Commit Example State L1 Cache Active Active Cache Line R W Begin_Transaction Log Buffer A = 15 1 A Write (A=15) Read (B) Persistent Memory Write (B=25) Transaction Log End_Transaction In-place Values A = 10 A = 10 A = 15 B = 20 B = 20 B = 25 C = 30 � 11

DHTM: Durable Hardware Transactional Memory Arpit Joshi , Vijay - PowerPoint PPT Presentation

DHTM: Durable Hardware Transactional Memory Arpit Joshi , Vijay Nagarajan, Marcelo Cintra, Stratis Viglas ISCA 2018 Persistent Memory is here 2 Persistent Memory is here 2 Persistent Memory Systems L1 L1 LLC Persistent Memory

Durable Transactional Memory Can Scale With TimeStone * R. Madhava Krishnan , Jaeho Kim * , Ajit

Hardware Transactional Memory Shao-Hung Chiu, Upasana Sridhar Transactional Memory - Where did

Extending Hardware Transactional Memory to Support Non-busy Waiting and Non-transactional Actions

NON-BLOCKING DATA STRUCTURES AND TRANSACTIONAL MEMORY Tim Harris, 27 November 2015 Lecture 8

NON-BLOCKING DATA STRUCTURES AND TRANSACTIONAL MEMORY Tim Harris, 25 November 2016 Lecture 8

Extending Hardware Transactional Memory Capacity via Rollback-Only Transactions and Suspend/Resume

A Comprehensive Study of Con fl ict Resolution Policies in Hardware Transactional Memory Ege

Hardware Transactional Memory on Haswell EP Viktor Leis Technische Universitt Mnchen 1 / 14

Taming Transactions: Towards Hardware-Assisted Control Flow Integrity using Transactional Memory

Hardware Acceleration of Transactional Memory on Commodity Systems Jared Casper , Tayo Oguntebi,

Hardware Acceleration of Transactional Memory on Commodity Systems Jared Casper , Tayo Oguntebi,

Mutex Locking versus Hardware Transactional Memory: An Experimental Evaluation Thesis Defense

Overview Limitations of lock-based programming Transactional memory Programming with

Lazy Hardware Transactional Memory Anurag Negi *, Rubn Titos-Gil^, Manuel E. Acacio^, Jose M.

6 Transactional Memory Chip Multiprocessors (ACS MPhil) Robert Mullins Overview

Transactional Memory 1 To read more This days papers: Herlihy and Moss, Transactional

SEER PROBABILISTIC SCHEDULING FOR COMMODITY HARDWARE TRANSACTIONAL MEMORY Nuno Diegues , Paolo

Evaluating the Impact of Transactional Characteristics on the Performance of Transactional Memory

Kathleen Fisher Reading: Beautiful Concurrency, The Transactional Memory / Garbage

VMM Emulation of Intel Hardware Transactional Memory Maciej Swiech, Kyle Hale, Peter Dinda

Locality-Adaptive Parallel Hash Joins using Hardware Transactional Memory ANIL SHANBHAG , HOLGER

Lowering the Overhead of Nonblocking Software Transactional Memory Virendra J. Marathe Michael

Window-Based Greedy Contention Management for Transactional Memory Gokarna Sharma ( LSU ) Brett

VIRTUES AND LIMITATIONS OF COMMODITY HARDWARE TRANSACTIONAL MEMORY Nuno Diegues, Paolo Romano and