Embed Size (px)
Citation preview
An Innovative Approach to Achieve CompositionalityEfficiently using Multi-Version Object Based
Transactional Systems
Chirag Juyal1 Sandeep Kulkarni2 Sweta Kumari1 Sathya Peri1
Archit Somani1
1CSE Dept, IIT Hyderabad 2CSE Dept, Michigan State University20th International Symposium on Stabilization, Safety, and
Security of Distributed Systems (SSS 2018)
IIT Hyderbad, INDIA MV-OSTMs 1 / 43
Outline
1 Introduction to STMs
2 Problem with read-write STMs
3 Object Based STMs
4 Motivation towards MV-OSTM
5 Experimental Evaluation
6 Conclusion
7 Future Work
IIT Hyderbad, INDIA MV-OSTMs 2 / 43
Outline
1 Introduction to STMs
2 Problem with read-write STMs
3 Object Based STMs
4 Motivation towards MV-OSTM
5 Experimental Evaluation
6 Conclusion
7 Future Work
IIT Hyderbad, INDIA MV-OSTMs 3 / 43
Software Transaction Memory Systems (STMs)Introduction
Transaction
Sequence of instructions guaranteed to execute atomically.
History
Concurrent execution of transactions.
w2(x , 10) w2(y , 10)
r1(x , 0)
T2
T1
C2
r1(z , 0)C1
Figure: History of an STM
IIT Hyderbad, INDIA MV-OSTMs 4 / 43
Software Transaction Memory Systems (STMs)Introduction
Transaction
Sequence of instructions guaranteed to execute atomically.
History
Concurrent execution of transactions.
w2(x , 10) w2(y , 10)
r1(x , 0)
T2
T1
C2
r1(z , 0)C1
Figure: History of an STM
IIT Hyderbad, INDIA MV-OSTMs 4 / 43
STMs Cont’d..Introduction
Software Transaction Memory Systems(STMs) are a convenientprogramming interface for a programmer to access shared memoryusing concurrent threads without worrying about concurrency issues.
Traditionally, STMs export the following methods:
t begin()t read()t write()tryC() and tryA()
We refer to these as Read-Write STMs(or RWSTM).
IIT Hyderbad, INDIA MV-OSTMs 5 / 43
STMs Cont’d..Introduction
Software Transaction Memory Systems(STMs) are a convenientprogramming interface for a programmer to access shared memoryusing concurrent threads without worrying about concurrency issues.
Traditionally, STMs export the following methods:
t begin()t read()t write()tryC() and tryA()
We refer to these as Read-Write STMs(or RWSTM).
IIT Hyderbad, INDIA MV-OSTMs 5 / 43
A thread safe concurrent linked-list(LL) using STM
Algorithm 1 Insert(LL, e): Invoked by a thread to insert an element e intoa linked-list LL. This method is implemented using transactions.
1: retry = 0;2: while (true) do3: id = t begin(retry);4: ...5: ...6: v = t read(id , x); . reads the value of x as v7: ...8: ...9: t write(id , x , v ′); . writes a value v ′ to x10: ...11: ...12: ret = tryC(id); . tryC can return commit or abort13: if (ret == commit) then14: break;15: else16: retry++;17: end if18: end while
IIT Hyderbad, INDIA MV-OSTMs 6 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Working of STMs methods
T2T1
T1’s Local Log T2’s Local Log
Memory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T1’s Local Log T2’s Local Log
Memory
Read fromMemory
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
T2’s Local Log
Memory
T1’s Local Log
Write to local Log
T2T1
Write to local Log
T2’s Local Log
Memory
T1’s Local Log
T2T1
Memory
T1’s Local Log T2’s Local Log
T2
Validate T1’s Log
T1
Memory
T2’s Local Log
T2T1
T1’s Log Permanent
Memory
T2’s Local Log
T1 Committed
T1
T1’s Log Permanent T2’s Log Validated
T2
Memory
T1 Committed
T1
T1’s Log Permanent
T2
Memory
T1 Committed T2 Aborted
Figure: Working of STM System
IIT Hyderbad, INDIA MV-OSTMs 7 / 43
Correctness of STM System: Opacity a
Opacity
1 It is a popular correctness-criteria for STMs which is stronger thanserializability.
2 Opacity like serializability requires that the concurrent executionincluding the aborted transactions be equivalent to some serialexecution.
aGuerraoui, R., Kapalka, M.: On the Correctness of Transactional Memory. PPoPP, 2008
IIT Hyderbad, INDIA MV-OSTMs 8 / 43
Correctness of STM System: Opacity a
Opacity
1 It is a popular correctness-criteria for STMs which is stronger thanserializability.
2 Opacity like serializability requires that the concurrent executionincluding the aborted transactions be equivalent to some serialexecution.
aGuerraoui, R., Kapalka, M.: On the Correctness of Transactional Memory. PPoPP, 2008
IIT Hyderbad, INDIA MV-OSTMs 8 / 43
Correctness of STM System Cont’d..
Example of opacity
H: r1(x,0)w2(x,10)w2(y,10)C2r1(y,A)A1
w2(x , 10) w2(y , 10)
A1r1(y , 10)r1(x , 0)
C2
T1
T2
Figure: History H is not Opaque but Serializable
w2(x , 10) w2(y , 10)
A1r1(y ,A)r1(x , 0)
T2
T1
C2
Figure: Opaque History H
IIT Hyderbad, INDIA MV-OSTMs 9 / 43
Correctness of STM System Cont’d..
Example of opacity
H: r1(x,0)w2(x,10)w2(y,10)C2r1(y,A)A1
w2(x , 10) w2(y , 10)
A1r1(y , 10)r1(x , 0)
C2
T1
T2
Figure: History H is not Opaque but Serializable
w2(x , 10) w2(y , 10)
A1r1(y ,A)r1(x , 0)
T2
T1
C2
Figure: Opaque History HIIT Hyderbad, INDIA MV-OSTMs 9 / 43
Correctness of STM System Cont’d..
Example of opacity
H: r1(x,0)w2(x,10)w2(y,10)C2r1(y,A)A1
w2(x , 10) w2(y , 10)
A1r1(y ,A)r1(x , 0)
T2
T1
C2
Figure: Opaque History H
A1r1(y ,A)r1(x , 0)T1
w2(x , 10) w2(y , 10)C2
T2
Figure: Equivalent serial history S: T1, T2
IIT Hyderbad, INDIA MV-OSTMs 10 / 43
Correctness of STM System Cont’d..
Example of opacity
H: r1(x,0)w2(x,10)w2(y,10)C2r1(y,A)A1
w2(x , 10) w2(y , 10)
A1r1(y ,A)r1(x , 0)
T2
T1
C2
Figure: Opaque History H
A1r1(y ,A)r1(x , 0)T1
w2(x , 10) w2(y , 10)C2
T2
Figure: Equivalent serial history S: T1, T2
IIT Hyderbad, INDIA MV-OSTMs 10 / 43
Outline
1 Introduction to STMs
2 Problem with read-write STMs
3 Object Based STMs
4 Motivation towards MV-OSTM
5 Experimental Evaluation
6 Conclusion
7 Future Work
IIT Hyderbad, INDIA MV-OSTMs 11 / 43
Problem with read-write STMs
k2 k5 k7 k8 k9
Figure: A sample concurrent list
T1: lookup(k5), lookup(k8) & T2: delete(k7)
r1(k5) w2(k5) r1(k5)w2(k7)
T2
r2(k2) r2(k5) r1(k8)r1(k2)
T1
r1(k2) r2(k7)
Figure: Tree Structure : conflicts are (r1(k5), w2(k5)) and (w2(k5), r1(k5))
IIT Hyderbad, INDIA MV-OSTMs 12 / 43
Problem with read-write STMs
k2 k5 k7 k8 k9
Figure: A sample concurrent list
T1: lookup(k5), lookup(k8) & T2: delete(k7)
r1(k5) w2(k5) r1(k5)w2(k7)
T2
r2(k2) r2(k5) r1(k8)r1(k2)
T1
r1(k2) r2(k7)
Figure: Tree Structure : conflicts are (r1(k5), w2(k5)) and (w2(k5), r1(k5))
IIT Hyderbad, INDIA MV-OSTMs 12 / 43
Problem with read-write STMs
k2 k5 k7 k8 k9
Figure: A sample concurrent list
T1: lookup(k5), lookup(k8) & T2: delete(k7)
r1(k5) w2(k5) r1(k5)w2(k7)
T2
r2(k2) r2(k5) r1(k8)r1(k2)
T1
r1(k2) r2(k7)
Figure: Tree Structure : conflicts are (r1(k5), w2(k5)) and (w2(k5), r1(k5))
IIT Hyderbad, INDIA MV-OSTMs 12 / 43
Problem at read-write level
r1(k5) w2(k5) r1(k5)w2(k7)
T2
r2(k2) r2(k5) r2(k7)r1(k2) r1(k8)r1(k2)
T1
Figure: Tree Structure
Schedule cannot be accepted by a RWSTM.
T2T1
Figure: Cycle (Not Serial)
IIT Hyderbad, INDIA MV-OSTMs 13 / 43
Problem at read-write level
r1(k5) w2(k5) r1(k5)w2(k7)
T2
r2(k2) r2(k5) r2(k7)r1(k2) r1(k8)r1(k2)
T1
Figure: Tree Structure
Schedule cannot be accepted by a RWSTM.
T2T1
Figure: Cycle (Not Serial)IIT Hyderbad, INDIA MV-OSTMs 13 / 43
Outline
1 Introduction to STMs
2 Problem with read-write STMs
3 Object Based STMs
4 Motivation towards MV-OSTM
5 Experimental Evaluation
6 Conclusion
7 Future Work
IIT Hyderbad, INDIA MV-OSTMs 14 / 43
OSTM b c
Introduction
Object-based STMs (OSTM) operate on higher level objects ratherthan primitive read & writes which act upon memory locations.
OSTM for sets may export t begin(), t insert(), t delete(), t lookup()and tryC().
OSTM for stacks may export t begin(), t push(), t pop(), t peek()and tryC().
bHerlihy, M., Koskinen, E., Transactional boosting: a methodology for highly-concurrent transactional objects. PPoPP’08
cHassan, Ahmed and Palmieri, Roberto and Ravindran, Binoy, Optimistic Transactional Boosting, PPoPP’14
IIT Hyderbad, INDIA MV-OSTMs 15 / 43
OSTM b c
Introduction
Object-based STMs (OSTM) operate on higher level objects ratherthan primitive read & writes which act upon memory locations.
OSTM for sets may export t begin(), t insert(), t delete(), t lookup()and tryC().
OSTM for stacks may export t begin(), t push(), t pop(), t peek()and tryC().
bHerlihy, M., Koskinen, E., Transactional boosting: a methodology for highly-concurrent transactional objects. PPoPP’08
cHassan, Ahmed and Palmieri, Roberto and Ravindran, Binoy, Optimistic Transactional Boosting, PPoPP’14
IIT Hyderbad, INDIA MV-OSTMs 15 / 43
OSTM b c
Introduction
Object-based STMs (OSTM) operate on higher level objects ratherthan primitive read & writes which act upon memory locations.
OSTM for sets may export t begin(), t insert(), t delete(), t lookup()and tryC().
OSTM for stacks may export t begin(), t push(), t pop(), t peek()and tryC().
bHerlihy, M., Koskinen, E., Transactional boosting: a methodology for highly-concurrent transactional objects. PPoPP’08
cHassan, Ahmed and Palmieri, Roberto and Ravindran, Binoy, Optimistic Transactional Boosting, PPoPP’14
IIT Hyderbad, INDIA MV-OSTMs 15 / 43
OSTMExecution at layer-1
k2 k5 k7 k8 k9
Figure: A sample concurrent list
T1: lookup(k5), lookup(k8) & T2: delete(k7)
Layer-1: Lookups & Deletes
Layer-0: Reads
r2(k2) r2(k5) r2(k7) w2(k7) r1(k2)
d2(k7)
T2
c2
r1(k5) r1(k5)w2(k5)
l1(k5) l1(k8)
r1(k2) r1(k8)
T1
& Writes
Figure: Tree Structure : no conflict at Layer-1 d
dGerhard Weikum and Gottfried Vossen. 2001. Transactional Information Systems Book
IIT Hyderbad, INDIA MV-OSTMs 16 / 43
OSTMExecution at layer-1
k2 k5 k7 k8 k9
Figure: A sample concurrent list
T1: lookup(k5), lookup(k8) & T2: delete(k7)
Layer-1: Lookups & Deletes
Layer-0: Reads
r2(k2) r2(k5) r2(k7) w2(k7) r1(k2)
d2(k7)
T2
c2
r1(k5) r1(k5)w2(k5)
l1(k5) l1(k8)
r1(k2) r1(k8)
T1
& Writes
Figure: Tree Structure : no conflict at Layer-1 d
dGerhard Weikum and Gottfried Vossen. 2001. Transactional Information Systems Book
IIT Hyderbad, INDIA MV-OSTMs 16 / 43
OSTMExecution at layer-1
k2 k5 k7 k8 k9
Figure: A sample concurrent list
T1: lookup(k5), lookup(k8) & T2: delete(k7)
Layer-1: Lookups & Deletes
Layer-0: Reads
r2(k2) r2(k5) r2(k7) w2(k7) r1(k2)
d2(k7)
T2
c2
r1(k5) r1(k5)w2(k5)
l1(k5) l1(k8)
r1(k2) r1(k8)
T1
& Writes
Figure: Tree Structure : no conflict at Layer-1 d
dGerhard Weikum and Gottfried Vossen. 2001. Transactional Information Systems Book
IIT Hyderbad, INDIA MV-OSTMs 16 / 43
OSTMExecution at layer-1
d2(k7)
T2
l1(k5) l1(k8)
T1
Figure: Pruned Tree
l1(k8) d2(k7)
T2
l1(k5)
T1
Figure: Sequential Schedule
T2T1
Figure: Serial History
IIT Hyderbad, INDIA MV-OSTMs 17 / 43
OSTMExecution at layer-1
d2(k7)
T2
l1(k5) l1(k8)
T1
Figure: Pruned Tree
l1(k8) d2(k7)
T2
l1(k5)
T1
Figure: Sequential Schedule
T2T1
Figure: Serial History
IIT Hyderbad, INDIA MV-OSTMs 17 / 43
OSTMExecution at layer-1
d2(k7)
T2
l1(k5) l1(k8)
T1
Figure: Pruned Tree
l1(k8) d2(k7)
T2
l1(k5)
T1
Figure: Sequential Schedule
T2T1
Figure: Serial HistoryIIT Hyderbad, INDIA MV-OSTMs 17 / 43
Inefficiency with OSTM
T1
T2
A1
C2
Lu1(ht, k2, nil)
Ins2(ht, k2, v2)
Lu1(ht, k1,Abort)
Ins2(ht, k1, v2)
Figure: Single version OSTM
IIT Hyderbad, INDIA MV-OSTMs 18 / 43
Outline
1 Introduction to STMs
2 Problem with read-write STMs
3 Object Based STMs
4 Motivation towards MV-OSTM
5 Experimental Evaluation
6 Conclusion
7 Future Work
IIT Hyderbad, INDIA MV-OSTMs 19 / 43
Motivation towards MV-OSTMAdvantages of multi-versione over single version RWSTM
T1
T2
A1
C2
r1(k2, 0)
w2(k2, 10) w2(k1, 20)
r1(k1,Abort)
Figure: Single version RWSTM
T1
T2
C2
r1(k2, 0)
w2(k2, 10) w2(k1, 20)
C1r1(k1, 0)
Figure: Multi-version RWSTM (MV-RWSTM) : (T1, T2)
eKumar P., Peri S., and K. Vidyasankar. A TimeStamp Based Multi-version STM Algorithm. ICDCN, 2014
IIT Hyderbad, INDIA MV-OSTMs 20 / 43
Motivation towards MV-OSTMAdvantages of multi-versione over single version RWSTM
T1
T2
A1
C2
r1(k2, 0)
w2(k2, 10) w2(k1, 20)
r1(k1,Abort)
Figure: Single version RWSTM
T1
T2
C2
r1(k2, 0)
w2(k2, 10) w2(k1, 20)
C1r1(k1, 0)
Figure: Multi-version RWSTM (MV-RWSTM) : (T1, T2)
eKumar P., Peri S., and K. Vidyasankar. A TimeStamp Based Multi-version STM Algorithm. ICDCN, 2014
IIT Hyderbad, INDIA MV-OSTMs 20 / 43
Proposed Algorithm : MV-OSTMAdvantages of multi-version over single version OSTM
T1
T2
A1
C2
Lu1(ht, k2, nil)
Ins2(ht, k2, v2)
Lu1(ht, k1,Abort)
Ins2(ht, k1, v2)
Figure: Single version OSTM
T1
T2
C2
C1Lu1(ht, k2, nil)
Ins2(ht, k2, v2)
Lu1(ht, k1, nil)
Ins2(ht, k1, v2)
Figure: Multi-version OSTM (MV-OSTM) : (T1, T2)
IIT Hyderbad, INDIA MV-OSTMs 21 / 43
Proposed Algorithm : MV-OSTMAdvantages of multi-version over single version OSTM
T1
T2
A1
C2
Lu1(ht, k2, nil)
Ins2(ht, k2, v2)
Lu1(ht, k1,Abort)
Ins2(ht, k1, v2)
Figure: Single version OSTM
T1
T2
C2
C1Lu1(ht, k2, nil)
Ins2(ht, k2, v2)
Lu1(ht, k1, nil)
Ins2(ht, k1, v2)
Figure: Multi-version OSTM (MV-OSTM) : (T1, T2)
IIT Hyderbad, INDIA MV-OSTMs 21 / 43
Proposed Algorithm : MV-OSTMHigh-level data structure
A transaction Ti is assigned a unique timestamp i , when invoked.
Timestamps are monotonically increasing.
Ti can issue t lookup(), t insert(), t delete() and tryC() methods.
k1
1
2
3
4
m
k3 k6
+∞−∞ k8k7
Figure: MV-OSTM designf
fSathya Peri, Ajay Singh, and Archit Somani, Efficient means of Achieving Composability using Object based Conflicts on
Transactional Memory, NETYS’18
IIT Hyderbad, INDIA MV-OSTMs 22 / 43
Proposed Algorithm : MV-OSTMHigh-level data structure
A transaction Ti is assigned a unique timestamp i , when invoked.
Timestamps are monotonically increasing.
Ti can issue t lookup(), t insert(), t delete() and tryC() methods.
k1
1
2
3
4
m
k3 k6
+∞−∞ k8k7
Figure: MV-OSTM designf
fSathya Peri, Ajay Singh, and Archit Somani, Efficient means of Achieving Composability using Object based Conflicts on
Transactional Memory, NETYS’18
IIT Hyderbad, INDIA MV-OSTMs 22 / 43
Proposed Algorithm : MV-OSTMHigh-level data structure cont’d..
RVL(Return Value List)
VL(Version List)k1
0 v0 rvl1T
201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23
rvl3
35
ts val rvlMark
Figure: Data Structure for maintaining versions
IIT Hyderbad, INDIA MV-OSTMs 23 / 43
Proposed Algorithm : MV-OSTMt lookup() method
t lookup() rule: Ti on invoking Lui (k1) lookups the value v insertedby a transaction Tj that commits before Lui (k1) and j is the largesttimestamp ≤ i .
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23 35
rvl3
13
Figure: Lookup on key k1 by T13
IIT Hyderbad, INDIA MV-OSTMs 24 / 43
Proposed Algorithm : MV-OSTMt lookup() method
t lookup() rule: Ti on invoking Lui (k1) lookups the value v insertedby a transaction Tj that commits before Lui (k1) and j is the largesttimestamp ≤ i .
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23 35
rvl3
13
Figure: Lookup on key k1 by T13
IIT Hyderbad, INDIA MV-OSTMs 24 / 43
Proposed Algorithm : MV-OSTMt lookup() method cont’d..
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23 35
rvl3
13
Figure: T13 searching appropriate place in version list of k1
IIT Hyderbad, INDIA MV-OSTMs 25 / 43
Proposed Algorithm : MV-OSTMt lookup() method cont’d..
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23 35
rvl3
13
Figure: T13 successfully added into rvl1
IIT Hyderbad, INDIA MV-OSTMs 26 / 43
Proposed Algorithm : MV-OSTMt insert(), t delete(), and tryC() methods
t insert() and t delete() rule: Ti inserts and delete into localmemory. The actual effect of both the methods will come in tryC().
tryC() rule: Ti on invoking tryC() operation checks for each key k ,in its Insert set and Delete set:
1 If a transaction Tk has lookups k from Tj and k is committed with j <i < k, then tryCi () returns abort.
2 otherwise, the transaction Ti is allowed to commit.
IIT Hyderbad, INDIA MV-OSTMs 27 / 43
Proposed Algorithm : MV-OSTMtryC() : t insert() method
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
13 201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23 35
rvl3
40
Figure: Insert a version of key k1 by T40
IIT Hyderbad, INDIA MV-OSTMs 28 / 43
Proposed Algorithm : MV-OSTMtryC() : t insert() method cont’d..
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
13 201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23 35
rvl3
40
Figure: T40 searching appropriate place in version list of k1
IIT Hyderbad, INDIA MV-OSTMs 29 / 43
Proposed Algorithm : MV-OSTMtryC() : t insert() method cont’d..
k1
0 v0 rvl1T
13 201675 2827 30
15 v5 F rvl2
12 18 23 35
40 F rvl3v1525 F rvl3v10
Figure: T40 successfully created a new version of k1
IIT Hyderbad, INDIA MV-OSTMs 30 / 43
Proposed Algorithm : MV-OSTMtryC() : t insert() method cont’d..
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
13 201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23
rvl3
40
45c
Figure: Insert a version of key k1 by T40
IIT Hyderbad, INDIA MV-OSTMs 31 / 43
Proposed Algorithm : MV-OSTMtryC() : t insert() method cont’d..
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
13 201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23
rvl3
40
45c
Figure: T40 searching appropriate place in version list of k1
IIT Hyderbad, INDIA MV-OSTMs 32 / 43
Proposed Algorithm : MV-OSTMtryC() : t insert() method cont’d..
VL(Version List)
RVL(Return Value List)
k1
0 v0 rvl1T
13 201675 2827 30
15 v5 F rvl2 25 v10 F
12 18 23
rvl3
40
45c
Figure: Abort T40 : T45 committed before T40
IIT Hyderbad, INDIA MV-OSTMs 33 / 43
Correctness of MV-OSTM g
Theorem (1)
Any history generated by HT-MVOSTM is opaque.
Theorem (2)
HT-MVOSTM with unbounded versions ensures that lookup methods donot return abort.
gArxiv Link: https://arxiv.org/abs/1712.09803
IIT Hyderbad, INDIA MV-OSTMs 34 / 43
Outline
1 Introduction to STMs
2 Problem with read-write STMs
3 Object Based STMs
4 Motivation towards MV-OSTM
5 Experimental Evaluation
6 Conclusion
7 Future Work
IIT Hyderbad, INDIA MV-OSTMs 35 / 43
Experimental Evaluation
Setup
Intel(R) Xeon(R) CPU, 32 GB RAM, 56 Threads.
We have compared proposed HT-MVOSTM with HT-OSTM / ElasticSTM(ESTM)a / Read Write STM (RWSTM) / Multi-Version Timestamp ordering Protocol(HT-MVTO).
We have compared proposed list-MVOSTM with list-OSTM /Boosted-list / NOrec-listb / Trans-listc / list-MVTO
aFelber, P., Gramoli, V., Guerraoui, R.: Elastic Transactions. J. Parallel Distrib. Comput.100(C), 2017
bDalessandro, L., Spear, M.F., Scott, M.L.: NOrec: Streamlining STM by Abolishing Own-ership Records. In
Govindarajan, R., Padua, D.A., Hall, M.W., eds.: PPOPP’10c
Zhang, D., Dechev, D.: Lock-free Transactions Without Rollbacks for Linked Data Structures. SPAA’16
IIT Hyderbad, INDIA MV-OSTMs 36 / 43
Experimental Evaluation Cont’d..
Lookup Intensive Workload: lookup:90%, insert:8% & delete:2%Update Intensive Workload: lookup:10%, insert:45% & delete:45%
2 4 8 1 6 3 2 6 40 . 0 0 0
0 . 0 0 2
0 . 0 0 4
0 . 0 0 6
0 . 0 0 8
0 . 0 1 0
2 4 8 1 6 3 2 6 40 . 0 0
0 . 0 5
0 . 1 0
0 . 1 5
0 . 2 0
0 . 2 5
0 . 3 0
( b ) U p d a t e I n t e n s i v e W o r k l o a d( a ) L o o k u p I n t e n s i v e W o r k l o a d
Time (
Sec.)
# o f t h r e a d s
H T - M V O S T M H T - O S T M E S T M R W S T M H T - M V T O
# o f t h r e a d s
Figure: Performance of HT-MVOSTM
Proposed HT-MVOSTM gives better performance while improving the concurrency.
IIT Hyderbad, INDIA MV-OSTMs 37 / 43
Experimental Evaluation Cont’d..
2 4 8 1 6 3 2 6 40 . 0 00 . 0 10 . 0 20 . 0 30 . 0 40 . 0 50 . 0 60 . 0 70 . 0 8
2 4 8 1 6 3 2 6 40 . 0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
0 . 6
( b ) U p d a t e I n t e n s i v e W o r k l o a d( a ) L o o k u p I n t e n s i v e W o r k l o a d
Time (
Sec.)
# o f t h r e a d s
l i s t - M V O S T M l i s t - O S T M T r a n s - l i s t B o o s t i n g - l i s t N O r e c - l i s t l i s t - M V T O
# o f t h r e a d s
Figure: Performance of list-MVOSTM
Proposed list-MVOSTM gives better performance while improving the concurrency.
IIT Hyderbad, INDIA MV-OSTMs 38 / 43
Outline
1 Introduction to STMs
2 Problem with read-write STMs
3 Object Based STMs
4 Motivation towards MV-OSTM
5 Experimental Evaluation
6 Conclusion
7 Future Work
IIT Hyderbad, INDIA MV-OSTMs 39 / 43
Conclusion
We have combined both multi-version and OSTM ideas carefullycalled as MV-OSTM, for harnessing greater concurrency in STMs.
HT-MVOSTM shows average speedup of 6.3, 11.2, 4.7, 2.7 times forstate of the art HT-MVTO, RWSTM, ESTM & HT-OSTMrespectively.
The average speedup achieved by list-MVOSTM from state of the artlist-MVTO, NOrec-list, Boosting-list, Trans-list, list-OSTM are 91.5,29.5, 23, 24, 2.1 respectively.
IIT Hyderbad, INDIA MV-OSTMs 40 / 43
Conclusion
We have combined both multi-version and OSTM ideas carefullycalled as MV-OSTM, for harnessing greater concurrency in STMs.
HT-MVOSTM shows average speedup of 6.3, 11.2, 4.7, 2.7 times forstate of the art HT-MVTO, RWSTM, ESTM & HT-OSTMrespectively.
The average speedup achieved by list-MVOSTM from state of the artlist-MVTO, NOrec-list, Boosting-list, Trans-list, list-OSTM are 91.5,29.5, 23, 24, 2.1 respectively.
IIT Hyderbad, INDIA MV-OSTMs 40 / 43
Conclusion
We have combined both multi-version and OSTM ideas carefullycalled as MV-OSTM, for harnessing greater concurrency in STMs.
HT-MVOSTM shows average speedup of 6.3, 11.2, 4.7, 2.7 times forstate of the art HT-MVTO, RWSTM, ESTM & HT-OSTMrespectively.
The average speedup achieved by list-MVOSTM from state of the artlist-MVTO, NOrec-list, Boosting-list, Trans-list, list-OSTM are 91.5,29.5, 23, 24, 2.1 respectively.
IIT Hyderbad, INDIA MV-OSTMs 40 / 43
Outline
1 Introduction to STMs
2 Problem with read-write STMs
3 Object Based STMs
4 Motivation towards MV-OSTM
5 Experimental Evaluation
6 Conclusion
7 Future Work
IIT Hyderbad, INDIA MV-OSTMs 41 / 43
ProgressFuture Work
MV-OSTM model can be extended to Starvation-free multi-versionOSTM (SF-MVOSTM).
MV-OSTM can be enlarged to other data structures like Queue,Stack, Tree etc.
An interesting aspect is exploring Nestingh for MV-OSTM in whichone object-based transaction invokes other read-write transaction.
hNi, Yang and Menon, Vijay S. and Adl-Tabatabai, Ali-Reza and Hosking, Antony L. and Hudson, Richard L. and Moss, J.
Eliot B. and Saha, Bratin and Shpeisman, Tatiana, Open Nesting in Software Transactional Memory, PPoPP ’07
IIT Hyderbad, INDIA MV-OSTMs 42 / 43
ProgressFuture Work
MV-OSTM model can be extended to Starvation-free multi-versionOSTM (SF-MVOSTM).
MV-OSTM can be enlarged to other data structures like Queue,Stack, Tree etc.
An interesting aspect is exploring Nestingh for MV-OSTM in whichone object-based transaction invokes other read-write transaction.
hNi, Yang and Menon, Vijay S. and Adl-Tabatabai, Ali-Reza and Hosking, Antony L. and Hudson, Richard L. and Moss, J.
Eliot B. and Saha, Bratin and Shpeisman, Tatiana, Open Nesting in Software Transactional Memory, PPoPP ’07
IIT Hyderbad, INDIA MV-OSTMs 42 / 43
ProgressFuture Work
MV-OSTM model can be extended to Starvation-free multi-versionOSTM (SF-MVOSTM).
MV-OSTM can be enlarged to other data structures like Queue,Stack, Tree etc.
An interesting aspect is exploring Nestingh for MV-OSTM in whichone object-based transaction invokes other read-write transaction.
hNi, Yang and Menon, Vijay S. and Adl-Tabatabai, Ali-Reza and Hosking, Antony L. and Hudson, Richard L. and Moss, J.
Eliot B. and Saha, Bratin and Shpeisman, Tatiana, Open Nesting in Software Transactional Memory, PPoPP ’07
IIT Hyderbad, INDIA MV-OSTMs 42 / 43
IIT Hyderbad, INDIA MV-OSTMs 43 / 43
