Options
Cloned Transactions : A New Execution Concept for Transactional Memory
Citation Link: https://doi.org/10.15480/882.1240
Publikationstyp
Doctoral Thesis
Date Issued
2015
Sprache
English
Author(s)
Advisor
Referee
Title Granting Institution
Technische Universität Hamburg
Place of Title Granting Institution
Hamburg
Examination Date
2015-04-01
Institut
TORE-DOI
Transactional memory aims to replace mutual exclusion in critical sections with
transactions on shared data to improve the scalability of concurrent
applications and eliminate traditional issues of parallel programming such as
deadlocks/livelocks. One key aspect of transaction processing is the concurrency
control, which is responsible to find an interleaving or overlapping schedule
for running transactions accessing the same shared data and provide a result
equal to some serial execution of the same transactions. In this regard, former
research focused on single-threaded concurrency control schemes using a trial and error
approach: The transaction repeats executing the critical section until
no conflict with other transactions was detected.
This thesis proposes the application of helper threads in transactional
memory to achieve a parallelised transaction processing, which is able to
compute the same critical section in respect to different serialisation orders with other transactions. While
traditional concurrency control has to rollback and start another attempt in the
same thread, this approach aims to compute the next attempt in a helper thread
in parallel to the current attempt computed by a leading thread. Due to the
nature of transactions and concurrent shared memory applications in general, the
introduction of helper threads comes with several new aspects to be considered
on the conceptional and the implementation level.
The work presents concepts for helper thread aided concurrency controls
independently executing the same transaction in parallel instead of sharing
information of the leader thread, to lower the contention. For example, the
detection of conflicts is delegated to a helper thread which executes the same
transaction on the same state of shared data in parallel instead of validating
the data read by the leader. In this manner parallelised transaction processing
has been proposed which executes a transaction in parallel with different
validation schemes: lazy and eager validation. Because the reaction time on
conflicts and execution time differ, they drift apart and compute results for different
serialisation orders.
The work also presents several requirements and sub-systems needed to run
parallelised transactions: The use of so-called transactional
wrappers or pure functions to speedup access to thread private data in
transactional sections cannot be granted, because the private data of the
application thread is now concurrently accessed by the helper thread. Also, a
concept has been developed to clone the state of the transaction start and
transfer the state of the helper committed transaction back to the leader. The
existing sandboxing approach of Dalessandro and Scott to suppress and
prevent errors in transactions resulting from lazy validation has been improved
and memory management inside transactions was adapted.
While the developed approaches can compete in some cases and show good
scalability especially in higher concurrency, the evaluation shows
an unexpected large loss of performance, too. An investigation of hardware
characteristics in respect to memory and cache read/write latency revealed
several scalability issues with concurrent access to the same address. These
issues have significant impact on highly concurrent applications,
which seriously affects the use of transactions in general and
parallelised transactions in particular.
transactions on shared data to improve the scalability of concurrent
applications and eliminate traditional issues of parallel programming such as
deadlocks/livelocks. One key aspect of transaction processing is the concurrency
control, which is responsible to find an interleaving or overlapping schedule
for running transactions accessing the same shared data and provide a result
equal to some serial execution of the same transactions. In this regard, former
research focused on single-threaded concurrency control schemes using a trial and error
approach: The transaction repeats executing the critical section until
no conflict with other transactions was detected.
This thesis proposes the application of helper threads in transactional
memory to achieve a parallelised transaction processing, which is able to
compute the same critical section in respect to different serialisation orders with other transactions. While
traditional concurrency control has to rollback and start another attempt in the
same thread, this approach aims to compute the next attempt in a helper thread
in parallel to the current attempt computed by a leading thread. Due to the
nature of transactions and concurrent shared memory applications in general, the
introduction of helper threads comes with several new aspects to be considered
on the conceptional and the implementation level.
The work presents concepts for helper thread aided concurrency controls
independently executing the same transaction in parallel instead of sharing
information of the leader thread, to lower the contention. For example, the
detection of conflicts is delegated to a helper thread which executes the same
transaction on the same state of shared data in parallel instead of validating
the data read by the leader. In this manner parallelised transaction processing
has been proposed which executes a transaction in parallel with different
validation schemes: lazy and eager validation. Because the reaction time on
conflicts and execution time differ, they drift apart and compute results for different
serialisation orders.
The work also presents several requirements and sub-systems needed to run
parallelised transactions: The use of so-called transactional
wrappers or pure functions to speedup access to thread private data in
transactional sections cannot be granted, because the private data of the
application thread is now concurrently accessed by the helper thread. Also, a
concept has been developed to clone the state of the transaction start and
transfer the state of the helper committed transaction back to the leader. The
existing sandboxing approach of Dalessandro and Scott to suppress and
prevent errors in transactions resulting from lazy validation has been improved
and memory management inside transactions was adapted.
While the developed approaches can compete in some cases and show good
scalability especially in higher concurrency, the evaluation shows
an unexpected large loss of performance, too. An investigation of hardware
characteristics in respect to memory and cache read/write latency revealed
several scalability issues with concurrent access to the same address. These
issues have significant impact on highly concurrent applications,
which seriously affects the use of transactions in general and
parallelised transactions in particular.
Subjects
Concurrent Programming
Transactional Memory
DDC Class
620: Ingenieurwissenschaften
Loading...
Name
machens.pdf
Size
1.92 MB
Format
Adobe PDF