stasis-aries-wal/doc/position-paper/LLADD.txt

Russell Sears
Eric Brewer
UC Berkeley

A Flexible, Extensible Transaction Framework

Existing transactional systems are designed to handle specific
workloads well.  Unfortunately, these systems' implementations are
mononolithic and hide the transactional infrastructure underneath a
SQL interface. Lower-level implementations such as Berkeley DB handle
a wider variety of workloads and are built in a more modular fashion.
However, they do not provide APIs to allow applications to build upon
and modify low-level policies such as allocation strategies, page
layout or details of recovery semantics.  Furthermore, data structure
implementations are typically not broken into separable, public APIs,
which discourages the implementation of new transactional data
structures.

Contrast this to the handling of data structures within modern
object-oriented programming languages such as C++ or Java.  Such
languages typically provide a large number of data storage algorithm
implementations.  These structures may be used interchangeably with
application-specific data collections, and collection implementations
may be composed into more sophisticated data structures.

We have implemented LLADD (/yad/), an extensible transactional storage
library that takes a composable and layered approach to transactional
storage.  Below, we present some of its high level features and
performance characteristics and discuss our plans to
extend the system into distributed domains.  Finally we introduce our
current research focus, the application of automated program
verification and optimization techniques to application specific
extensions.  Such techniques should significantly enhance the
usability and performance of our system, allowing application 
developers to implement sophisticated cross-layer optimizations easily.

Overview of the LLADD Architecture
----------------------------------

General-purpose transactional storage systems are extremely complex
and only handle certain types of workloads efficiently.  However, new
types of applications and workloads are introduced on a regular basis.
This results in the implementation of specialized, ad-hoc data storage
systems from scratch, wasting resources and preventing code reuse.

Instead of developing a set of general purpose data structures that
attempt to behave well across many workloads, we have implemented a
lower-level API that makes it easy for application designers to
implement specialized data structures.  Essentially, we have
implemented an extensible navigational database system.  We
believe that this system will support modern development practices and
allows transactions to be used in a wider range of applications.

Typically, implementations of general-purpose declarative systems 
are unable to keep up with the new classes of workloads introduced 
by rapidly evolving applications.  We believe that our architecture's
flexibility allows us to address such applications rapidly.  In cases
where the development of a general-purpose system is not economical, 
our system seems to be a reasonable long-term solution.  XML storage
technologies, which are rapidly evolving and still fail to handle 
many types of applications provide a good example.

For instance, most general-purpose solutions for semi-structured
information have difficulty handling computationally intensive
workloads posed by the large repositories that are typical of
bioinformatics research[PDB, NCBI, Gene Ontology].  These sytems are
slowly transitioning to XML, which currently has clear value as an
interchange format, but many of the data processing applications that
use these databases still employ ad-hoc solutions for data managment.
Whether or not general purpose XML database systems eventually meet
the unique needs of each of these scientific applications, extensions 
implemented on top of a more flexible data storage implementation 
could have avoided the need for ad-hoc solutions, and could serve 
as a partial prototype for higher level implementations.

LLADD is based upon an extensible version of ARIES but does not
hard-code details such as page format or data structure
implementation.  It provides a number of "operation" implementations
which consist of redo/undo methods and wrapper functions.  The redo/undo 
methods manipulate the page file by applying log entries while the 
wrapper functions produce log entries.  Redo methods handle all page
file manipulation during normal forward operation, reducing the amount 
of code that must be developed in order to implement new data structures.
LLADD handles the scheduling of redo/undo invocations, disk I/O, and all 
of the other details specified by the ARIES recovery algorithm, allowing
operation implementors to focus on the details that are important to the 
functionality their extension provides.

LLADD ships with a number of default data structures and
layouts, ranging from byte-level page layouts to linear hashtables 
and application-specific recovery schemes and data structures.  
These structures were developed with reusability in mind, encouraging
developers to compose existing operations into application-specific data
structures.  For example, the hashtable is 
implemented on top of reusable modules that implement a resizable array 
and two exchangeable linked-list variants.  

In other work, we show that the system is competitive with
Berkeley DB on traditional (hashtable based) workloads, and have shown
significant performance improvements for less conventional workloads
including custom data structure implementations, graph traversal
algorithms and transactional object persistence workloads.

The transactional object persistence system was based upon the
observation that most object perstistence schemes cache a second copy
of each in-memory object in a page file, and often keep a third copy
in operating system cache.  By implementing custom operations that
assume the program maintains a correctly implemented object cache, we
allow LLADD to service object update requests without updating the
page file.  

Since LLADD implements no-force, the only reason to update
the page file is to service future application read requests.
Therefore, we defer page file updates until the object is evicted from
the application's object cache, eliminating the need to maintain a large 
page cache in order to efficiently service write requests.  We also 
leveraged our customizable log format to log differences to objects 
instead of entire copies of objects.

With these optimizations, we showed a 2-3x performance improvement
over Berkeley DB on object persistence across our benchmarks, and a
3-4x improvement over an in-process version of MySQL with the InnoDB
backend.  (A traditional MySQL setup that made use of a separate
server process was prohibitively slow.  InnoDB provided the best
performance among MySQL's durable storage managers.)  Furthermore, our
system uses memory more efficiently, increasing its performance
advantage in situations where the size of system memory is a
bottleneck.

We leave systematic performance tuning of LLADD to future work, and
believe that further optimizations will improve our performance on
these benchmarks significantly.  In general, LLADD's customizability
enables many optimizations that are difficult for other systems.

Because of its natural integration into standard
system software development practices, we think that LLADD can be
naturally extended into networked and distributed domains.
For example, typical write-ahead-logging protocols implicitly
implement machine independent, reorderable log entries in order to
implement logical undo.  These two properties have been crucial in
past system software designs, including data replication,
distribution, and conflict resolution algorithms.  Therefore, we plan
to provide a networked, logical redo log as an application-level
primitive, and to explore system designs that leverage this approach.

Current Research Focus
----------------------

LLADD's design assumes that application developers will
implement high-performance transactional data structures.  This is a
big assumption, as these data structures are notoriously difficult to
implement correctly.  Our current research attempts to address these
concerns.

For our infrastructure to be generally useful the functionality that
it provides should be efficient, reliable and applicable to new
application domains.  We believe that improvements to the development
process can address each of these goals.

Application developers typically have a limited amount of time to
spend implementing and verifying application-specific storage
extensions, but bugs in these extensions have dire consequences.
Also, while data structure algorithms tend to be simple and easily
understood, performance tuning and verification of implementation
correctness is extremely difficult.

Recovery-based algorithms must behave correctly during forward
operation and also under arbitrary recovery scenarios.  The latter
requirement is particularly difficult to verify due to the large
number of materialized page file states that could occur after a
crash.

Fortunately, write-ahead-logging schemes such as ARIES make use of
nested-top-actions to vastly simplify the problem.  Given the
correctness of page-based physical undo and redo, logical undo may
assume that page spanning operations are applied to the data store
atomically.

Existing work in the static-analysis community has verified that
device driver implementations correctly adhere to complex operating
system kernel locking schemes[SLAM]. If we formalize LLADD's latching
and logging APIs, we believe that analyses such as these will be
directly applicable, allowing us to verify that data structure
behavior during recovery is equivalent to the behavior that would
result if an abort() was issued on each prefix of the log that is
generated during normal forward operation.

By using coarse latches that are held throughout entire logical 
operation invocations, we can
drastically reduce the size of this space, allowing conventional
state-state based search techniques (such as randomized or exhaustive
state-space searches, or unit testing techniques) to be
practical.  It has been shown that such coarse-grained latching can
yield high-performance concurrent data structures if
semantics-preserving optimizations such as page prefetching are
applied [ARIES/IM].

A separate approach to the static analysis of LLADD extensions uses
compiler optimization techniques.  Software built on top of layered
APIs frequently makes repeated calls to low level functions that result
in repeated work.  A common example in LLADD involves loops over data with
good locality in the page file.  The vast majority of the time, these
loops result in a series of high level API calls that repeatedly pin
and unpin the same underlying data.

The code for each of these high level API calls could be copied into
many different variants with different pinning/unpinning and
latching/unlatching behavior, but this would greatly complicate the
API that application developers must work with, and complicate any
application code that made use of such optimizations.

Compiler optimization techniques such as code hoisting and partial common 
subexpression elimination solve analogous problems to remove redundant 
computations.  Code hoisting moves code outside of loops and conditionals, while
partial common subexpression elimination inserts checks that decide at runtime
whether a particular computation is redundant.
We hope to extend such techniques to reduce the number
of buffer manager and locking calls made by existing code.  In 
situations where memory is abundant, these calls are a significant 
performance bottleneck, especially for read-only operations.

Anecdotal evidence and personal experience suggest that similar
optimization techniques are applicable to application code.  Because
local LLADD calls are simply normal function calls, it may even be
possible to apply the transformations that these optimizations perform
to application code that is unaware of the underlying storage
implementation.  This class of optimizations would be very difficult
to implement with existing transactional storage systems but should
significantly improve application performance.

We hope to validate our ideas about static analysis by incorporating
them into the development process as we increase the reliability and
overall quality of LLADD's implementation and its APIs.

Our architecture provides a set of tools that allow applications to implement
custom transactional data structures and page layouts.  This avoids
"impedance mismatch," simplifying applications and providing appropriate
applications with performance that is comparable or superior to other 
general-purpose solutions.  
By adding support for automated code verification and
transformations we hope to make it easy to produce correct extensions
and to allow simple, maintainable implementations to compete with
special purpose, hand-optimized code.