<?xml version="1.0" encoding="ISO-8859-1" ?>
<rss version="2.0">
 <channel>
  <title>URCS Seminars &amp; Talks</title>
  <link>http://www.cs.rochester.edu/dept/seminar</link>
  <description>Upcoming Seminars &amp; Talks from the Computer Science Department at the University of Rochester</description>
  <language>en-us</language>
  <image>
   <url>http://www.cs.rochester.edu/images/monalogo.jpg</url>
   <title>Computer Science Department, University of Rochester</title>
   <link>http://www.cs.rochester.edu/dept/seminar</link>
  </image>
  <lastBuildDate>Mon, 10 Nov 2008 10:57:24 -0500</lastBuildDate>
  
   <item>
    <title>Piotr Faliszewski: Manipulation of Elections: Algorithms and Infeasibility Results</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/365</link>
    <description>[Thursday, November 20, 2008 at  3:30 PM in Computer Studies  Building Room 601] Voting and elections are at the core of democratic societies. People vote to elect leaders,
decide policies, and organize their lives, but elections also have natural applications
in computer science. For example, agents in multiagent systems often need to work
together to complete some task, but each agent may have its own set of beliefs, preferences,
and goals. Voting provides agents with a natural way to reach decisions that
take all their preferences into account. With elections playing such an important role
both in real-life political settings and in computer science, it is natural to ask about
their resistance to misuse.

Two particular types of election misuse are manipulation and bribery. In manipulation,
a group of voters chooses to misrepresent its preferences in order to obtain a
more desirable outcome, and in bribery an outside agent, the briber, asks (possibly at
a cost) a group of voters to change its votes, to obtain some outcome desirable for the
briber. Classical results from political science show that, for any reasonable election
system, there are scenarios where at least some voters have an incentive to attempt
manipulation.

In this thesis we seek to protect elections from manipulators and bribers by making
their computational task of finding good manipulations/bribes prohibitively expensive.
When this is not possible, we seek to better understand (and even improve) the algorithmic
attacks that manipulators and bribers can employ. In doing so, we develop
new models of manipulation and bribery, and provide new approaches to studying the
computational complexity of bribery and manipulation in elections.</description>
    <pubDate>Mon, 10 Nov 2008 10:57:24 -0500</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/365</guid>
   </item>
  
   <item>
    <title>Christopher Stewart: Performance Modeling and System Management for Internet Services</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/364</link>
    <description>[Friday, October 24, 2008 at  2:00 PM in Computer Studies  Building Room 601] Internet services execute complex logic on heterogeneous and distributed server platforms for concurrent requests--- in short, they are complicated. Often subtle changes in request patterns or system configurations have unexpected and undesirable performance consequences. This dissertation demonstrates practical system management for Internet services that improves performance dependability, increases raw performance, and raises the performance-to-cost ratio in datacenters. In our approach, system management is guided by easy-to-understand and easy-to-apply models of a service&#39;s performance.  These performance models are easy to understand, because they use real system properties that are common across many services to characterize application-level performance (e.g., response time and throughput). They are easy to apply, because they are calibrated with system measurements that can be collected in practice. And they are useful in system management, because they predict what performance would be like if contemplated actions were taken. In particular, this dissertation presents case studies of our model-driven system management approach for actions frequently contemplated in practice, such as what if the component placement strategy were changed, what if the load balancing policy were changed, and what if services were consolidated. Experimental results with several Internet services under realistic operating conditions show the potential for model-driven system management.

</description>
    <pubDate>Tue, 14 Oct 2008 08:37:50 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/364</guid>
   </item>
  
   <item>
    <title>Chuanpeng Li: Operating System Enhancements for Data-Intensive Server Systems</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/343</link>
    <description>[Monday, August 11, 2008 at  2:00 PM in Computer Studies  Building Room 601] Recent studies on operating system support for concurrent server systems mostly target CPU-intensive workloads with light disk I/O activities. However, an important class of server systems that access a large amount of disk-resident data, such as the index searching server of large-scale Web search engines, has received limited attention. In this thesis work, we examine operating system techniques to improve the performance of data-intensive server systems under concurrent execution. We propose OS enhancements in three aspects of the operating system: file system prefetching, memory management, and disk I/O system. First, we propose a competitive prefetching strategy that can balance the overhead of disk I/O switching and the wasted I/O bandwidth on prefetching unnecessary data. Second, we explore a new memory management scheme for prefetched data, in order to reduce prefetching-incurred page thrashing at high execution concurrency. Third, we present five performance anomalies we identified in the Linux I/O system, and provide general discussions on operating system performance debugging.

We have implemented the proposed techniques in the Linux 2.6 kernel. Performance evaluation on microbenchmarks and real applications shows promising results.

</description>
    <pubDate>Tue, 29 Jul 2008 14:29:47 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/343</guid>
   </item>
  
   <item>
    <title>Virendra Marathe: Enhancing the Performance and Usability of Software Transactional Memory</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/344</link>
    <description>[Tuesday, August 12, 2008 at  2:00 PM in Computer Studies Building, Room 601] Transactional Memory (TM) is an emerging concurrency control technology
that promises to significantly ease the difficult task of programming on
multicore systems.  Foundational work on TM grew out of research in nonblocking concurrent algorithms.  Hence early software TM (STM) systems
were all nonblocking.  However they incurred high overheads, rendering them largely impractical.

More recently, there has been a shift toward blocking implementations of
STM.  This has been accompanied by several significant performance
optimizations, making blocking STMs far more practical.  A general hypothesis, based on informal arguments, has emerged that nonblocking STMs are fundamentally slower than blocking STMs.  In this talk I will show that this hypothesis is wrongly founded, and that nonblocking STMs can approach the performance of, and even outperform, state-of-the-art blocking STMs.  I will present several variants of our most recent nonblocking STM that are competitive with state-of-the-art blocking STMs.

</description>
    <pubDate>Tue, 29 Jul 2008 14:28:32 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/344</guid>
   </item>
  
   <item>
    <title>Ashwin Lall: Streaming Algorithms for Network Flow Data Analysis</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/342</link>
    <description>[Thursday, July 03, 2008 at 10:00 AM in Computer Studies Bldg. Room 601] The algorithmic field of Data Streaming was first proposed over a decade ago by Alon, Matias, and Szegedy. Since then, considerable work has been done in the streaming model of computation. The streaming model is precisely the one we want to use to perform real-time computation on network data because it allows only single pass access to the data and guarantees low memory consumption and processor usage per packet. However, the ideas from the streaming context have been slow to catch on in the networking literature. In this dissertation we will seek to catalyze this change by demonstrating both theoretically and empirically that the ideas from streaming algorithms will have a considerable impact on the performance of algorithms for online network analysis. In particular, we present novel algorithms for maintaining counts of large number of items and for computing the entropy of network traffic to demonstrate that streaming algorithms can solve these problems remarkably well in practice.

</description>
    <pubDate>Wed, 11 Jun 2008 09:11:36 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/342</guid>
   </item>
  
   <item>
    <title>Hao Zhang: Handling Complexity of Synchronous Grammars for Machine Translation</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/340</link>
    <description>[Tuesday, May 27, 2008 at  2:30 PM in Computer Studies Bldg. Room 601] Synchronous grammars are grammars that model two languages and their translational equivalence. They are rewriting systems extended to two dimensions. Systems based on synchronous grammars and tree transducers promise to improve the quality of statistical machine translation output, but are often very computationally intensive. We improve the efficiency of such systems, making both decoding and training fast and effective. 

We devise an algorithm for factoring syntactic re-orderings by binarizing synchronous rules when possible and show that the resulting rule set significantly improves the speed and accuracy of a state-of-the-art syntax-based machine translation system. 

We take a multi-pass approach to machine translation decoding when using synchronous context-free grammars as the translation model and n-gram language models: the first pass uses a bigram language model, and the resulting parse forest is used in the second pass to guide search with a trigram language model. An additional fast decoding pass maximizing the expected count of correct translation hypotheses increases the BLEU score significantly.

We combine the strengths of Bayesian modeling and synchronous grammar in unsupervised learning of basic translation phrase pairs. The structured space of a synchronous grammar is a natural fit for phrase pair probability estimation, though the search space can be prohibitively large. Therefore we explore efficient algorithms for pruning this space that lead to empirically effective results.

</description>
    <pubDate>Mon, 12 May 2008 08:37:24 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/340</guid>
   </item>
  
   <item>
    <title>Tongxin Bai: Blended Analysis for Speculative Specialization and Coarse-Grained Redundancy Detection</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/331</link>
    <description>[Monday, April 28, 2008 at  9:30 AM in Computer Studies Bldg. Room 632] Program specialization finds additional optimization opportunities by
targeting some rather than all executions of a program. The narrower
input domain increases the the number of runtime invariants and hence
exposes more runtime redundancies. Program specialization improves
program performance by eliminating the runtime redundancies. However,
the types of invariants that traditional specialization methods can
make use of are extremely limited, typically scalar variables. I
propose a set of new techniques, spanning from compiler and profile
analysis to specialization transformation, to extend the capability of
program specialization so that it can utilize the invariance hidden in
complex data structures. In the second half of the proposal I intend
to measure the similarity between runtime function instances. With
this information the analysis helps programmers understand program
performance issues in terms of computation redundancies. Besides the
redundancy issues memory bandwidth bottleneck is another performance
concern, particularly in this multi-core era. We propose a technique
called {\it partial caching of working sets} (PAW) to reduce cache
misses in a shared-cache environment through intentionally pinning
down a small part of each large working set in the cache.</description>
    <pubDate>Fri, 25 Apr 2008 11:44:38 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/331</guid>
   </item>
  
   <item>
    <title>Nicholas Morsillo: Semi-Supervised Probabilistic Models for Large-Scale Dataset Construction</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/336</link>
    <description>[Tuesday, April 29, 2008 at  1:30 PM in Computer Studies Bldg. Room 632] Object recognition is a fundamental problem for computer vision with a wide range of practical applications.  Although the field of visual recognition continues to hold great research promise, state of the art recognition systems have had limited real-world success.  A primary drawback of current methods is the inability to scale with large amounts of training data and many visual categories.

Scalability is increasingly crucial as the availability of digital imagery increases.  The internet hosts billions of images which can potentially serve as rich sources of training data for recognition.  Our work will explore this option through the following contributions.   1.  A new method for collecting large, clean image sets of specific object categories from web image search.  Our approach will use metadata including web page text to reduce the need for supervision, and it will exploit properties of web images for automatic segmentation and discovery of contextual information.  2. Creation of the first large-scale image dataset with clean labels, segmentation masks, and context descriptors.  3. A novel probabilistic model for visual recognition to efficiently combine object and contextual information from large-scale datasets.  We will thoroughly evaluate our approach by testing recognition performance on known datasets, using our processed web images for training.  </description>
    <pubDate>Thu, 24 Apr 2008 14:54:56 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/336</guid>
   </item>
  
   <item>
    <title>Xiao Zhang: On-chip Shared Resource Management in the Multicore Era</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/333</link>
    <description>[Tuesday, April 29, 2008 at 12:30 PM in Computer Studies Bldg. Room 703] Because of overwhelming heat concerns, processor manufacturers have shifted to the use of multiple cores on a single chip.
One of the key features of this new architecture is that all cores on a single chip share some cache (usually the last level cache) and
off-chip bandwidth.
This sharing presents new challenges due to the subtle interactions
of simultaneously executing processes.
For example, processes that are simultaneously accessing the shared cache may conflict with each other and result in skewed performance.

Today&#39;s operating systems are oblivious to to the needs of individual process for on-chip resources.
Multi-core operating systems must be made aware of on-chip resource competition in order for effective utilization of chip resources.
Using hardware performance counter information, it is possible to provide consolidated operating system support to efficiently manage
shared on-chip resources, particularly the shared cache and memory bus
bandwidth. I will explore both hardware-assisted and OS-level software techniques to improve the performance, fairness, and scalability of
multicore systems.</description>
    <pubDate>Thu, 24 Apr 2008 11:27:34 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/333</guid>
   </item>
  
   <item>
    <title>Shaojun Zhao: Applying Word Alignment to Automatic Machine Translation Evaluation</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/332</link>
    <description>[Monday, April 28, 2008 at  3:30 PM in Computer Studies Bldg. Room 703] All statistical machine translation (MT) systems are built on the idea of word
alignment. The HMM based word alignment model and the IBM model 4 play
the central role in word alignment. The HMM model has an efficient algorithm,
while the IBM model 4 has to perform sampling. We propose a model that has a
simple, exact dynamic programming algorithm, and achieves better performance
than the HMM model.
Automatic machine translation evaluation has revolutionized the development
of MT systems. The BLEU and NIST metric are the de-facto standards. They
are both based on N-gram overlap. Many other metrics have been proposed based
on the longest common subsequence. We propose several metrics
including a multidimensional alignment
approach which is likely to solve many of the drawbacks of the current metrics.</description>
    <pubDate>Thu, 24 Apr 2008 11:25:48 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/332</guid>
   </item>
  
   <item>
    <title>Ross Messing: Activity Recognition using Augmented Feature Flow</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/334</link>
    <description>[Tuesday, April 29, 2008 at  4:00 PM in Computer Studies Bldg. Room 703] Activity recognition systems have made extensive use of local motion,
but have largely failed to use long-range motion information.  We
develop a new way to represent the motion information in a video
sequence that captures long-range motion of features.  We show that
this information can be used to effectively disambiguate activities,
particularly when it is augmented by information orthogonal to motion,
like position.  We propose a number of ways to extend this work.  We
will investigate new representations of feature motion, and how to
tractably characterize the relationship between a pair of moving
features.  We will also explore powerful ways to augment the motion
information with much stronger position information (relative location
of features to a visual landmark) and appearance information (object
recognition on image patches extracted from the feature).  Lastly, we
propose a number of ways to make our model more robust and
nonparametric.</description>
    <pubDate>Thu, 24 Apr 2008 11:18:24 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/334</guid>
   </item>
  
   <item>
    <title>Hemayet Hossain: Effective on-Chip Cache Utilization in Chip Multiprocessors</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/335</link>
    <description>[Wednesday, April 30, 2008 at 10:00 AM in Computer Studies Bldg. Room 601] CMOS scaling trends allow increasing numbers  of transistors on a single chip. In
order to take advantage of these trends with manageable design complexity
while staying within power budgets, processor designers are increasingly turning
toward multi-core architectures --- often chip multiprocessors (CMPs) of
simultaneous multithreaded (SMT) cores --- for all application domains.

Increasing the number of cores on a single chip leads to higher demand on
on-chip cache capacity and interconnect bandwidth (on/off-chip). Scalable cache
coherence protocols support a large number of cores by reducing coherence
bandwidth requirements (e.g., directory). However, they squander away the low
latency cache-cache transfers (possible due to on-chip interconnects) by
introducing a level of indirection on the critical path of cache misses. Many
commercial, scientific, and data-mining workloads exhibit fine-grain (both
temporal and spatial) data-sharing patterns (due to communication and
synchronization) and desire low latency core-core data transfers but are penalized
by these scalable protocols. Furthermore, their working sets well exceed the
private cache sizes and stress-test the shared cache levels.

In this work, we (1) exploit low latency on chip interconnects and (2) improve
the utilization of shared caches. First, we super-impose a direct access
substrate on an existing scalable protocol to take advantage of low latency
interconnects. We show how this mechanism can be used to efficiently support
fine-grain sharing in applications. Additional benefits include reduced power
consumption and reduced interconnect bandwidth. We address location and
placement of data via novel uses of prediction and an adaptive replication,
migration, and producer-consumer optimization(ARMCO) cache coherence protocol.
Second, to meet the cache capacity demands of multiple cores and to improve the
utilization efficiency of on-chip cache resources we propose a sharing-aware
adaptive replacement policy (ShARP) and decoupling of tag and data. Finally, we
will analyze the effects of technology on, and scaling implications of, the
proposed protocol as part of this work.</description>
    <pubDate>Thu, 24 Apr 2008 10:46:46 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/335</guid>
   </item>
  
   <item>
    <title>Bin Wei: Towards Building Knowledge System from Text</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/328</link>
    <description>[Friday, April 25, 2008 at  1:30 PM in Computer Studies Bldg. Room 601] With growing amount of text data on line, especially some well-organized corpus like 
the Wikipedia, to make a better use of the data, (e.g: extract useful knowledge ) becomes 
a more and more interesting problem. This proposal proposes the thesis work on building 
knowledge system from normal, non-structured texts. The proposed approach will contain 
two steps: transform texts into relational database (this transformation should be able to work 
under semi-supervised manner)and build Markov Logic Network, a knowledge representation 
which combines probablistic and logic. 

</description>
    <pubDate>Mon, 21 Apr 2008 13:08:35 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/328</guid>
   </item>
  
   <item>
    <title>Paul Ardis: Video Inpainting Using Human Attention Modeling</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/329</link>
    <description>[Friday, April 25, 2008 at  3:30 PM in Computer Studies Bldg. Room 632] Inpainting techniques, for all that they have accomplished, continue to fail to satisfy.  By refusing to quantify result quality and approximating human visual expectation in an ad-hoc fashion, current methods remain innovationally stagnant.  Furthermore, modern methods that attempt to diverge from their traditional roots in data interpolation are faced with the risk of computational intractibility.  I propose to improve this situation over the next two or three years by making use of human attentional modeling and statistical learning techniques, ultimately providing a clear indication of advancement in the field. Specifically, I intend to:  1.) learn a decision boundary between patch-based methods and interpolation methods using a modified Conditional Random Field, 2.) compile a high-resolution video dataset for training and benchmark comparison of methods in the field, 3.) analyze the quality and efficiency of results produced using the learned decision boundary and simple existing methods as options for selection, 4.) adjust existing salience models and produce a novel method of computation learned from natural statistics in existing eyetrack datasets, 5.) produce a novel interpolation method based upon a salience-guided analysis of global and local image statistics, and 6.) compare my work against a number of leading algorithms in a series of double-blind human subject experiments as well as studying computational efficiency.

</description>
    <pubDate>Thu, 17 Apr 2008 11:17:59 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/329</guid>
   </item>
  
   <item>
    <title>Manu Chhabra: Optimality in Motor Control</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/305</link>
    <description>[Tuesday, February 12, 2008 at 11:00 AM in Computer Studies Bldg. Room 703] Several models of human motor control propose that stereotypical
movements like reaching are optimal, in the sense that they minimize
a performance cost function. However, from a computational point of view,
optimal control is a hard problem. First, most natural bio-mechanical systems
are redundant---there are more degrees of freedom than required. Second,
optimal control tasks often suffer from the ``curse of dimensionality&#39;&#39;. For example,
consider moving a two-joint arm from location A to location B in 100
time steps.  If torques are discretized to one of 10 values, then there are
10200 possible torque sequences. How do we efficiently search a space of that size?
In this talk, we address these issues from the perspective of
optimal control theory.

First, if optimal control is hard, how do humans accurately solve day to day
control problems? In the first part of the talk,  I propose that for
stereotypical movements like reaching, the optimal control solution can be decomposed as a sum of scaled
and time shifted components, or ``motor synergies&#39;&#39;. These synergies were
discovered through dimensionality reduction.
Near-optimal control is achieved by linearly combining the synergies.
Second, are humans optimal under non-stereotypical conditions?
In the second part of the talk, I present results from experiments that
show that humans are able to perform well even under certain non-stereotypical
 control regimes. In these experiments, subjects were required to
control a dynamical system corrupted with noise.
 A comparison of human performance to the theoretically
optimal solution shows that humans reach close to optimal performance under a
variety of noise conditions. Additionally, we showed that subjects
adapt to the noise regime rather than
using a fixed controller across different noise conditions.

</description>
    <pubDate>Wed, 30 Jan 2008 12:27:32 -0500</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/305</guid>
   </item>
  
   <item>
    <title>Benjamin Van Durme: Conditional Knowledge Acquisition</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/270</link>
    <description>[Friday, September 28, 2007 at 11:00 AM in Computer Science Bldg. Room 601] Research in Information Extraction has been focused on the extraction of
facts concerning individuals as compared to general knowledge pertaining to classes of
entities and events. Today&#39;s extraction systems are built assuming large, heterogeneous 
collections of text, and are based on simple techniques enabling high volume throughput.

We propose to extend the state of the art in Information Extraction by placing
greater focus on language understanding, the knowledge representation being targeted, 
and the underlying data upon which our system will run. We specifically target the
extraction of conditional relationships tending to hold between episodes.</description>
    <pubDate>Tue, 25 Sep 2007 11:19:50 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/270</guid>
   </item>
  
   <item>
    <title>Ming Zhong:  &quot; Exploiting Object Usage Nonuniformity in Distributed Data Management &quot;</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/209</link>
    <description>[Friday, July 06, 2007 at 10:00 AM in Computer Science Bldg. Room 601] Many emerging Internet services provide online, massive data service
backed by distributed system infrastructures. Examples include web
search engines, web-based email servers, media servers, online stores,
etc. Compared with traditional distributed systems, these distributed
data-intensive systems focus on data management and impose new challenges
in order to achieve fast, reliable, and massive information service to
many concurrent users.

It is well known that data objects hosted by Internet services often
exhibit highly skewed usage probability distributions.  Motivated by this,
we study the problem of exploiting data object usage popularity skewness
to improve the data management of common distributed data-intensive
applications. Our main results originate from the simple intuition that
being biased towards popular objects may increase system performance
compared with unbiased data management. However, choosing the optimal
biased strategy often amounts to a NP-hard problem. Furthermore,
incorporating the chosen popularity-driven strategy into an existing
application may demand significant system adjustment or rebuilding efforts in practice.

On the algorithmic side, we present results on deriving optimal/near-optimal
popularity-driven data management strategies for the applications we study.
Specifically, the results include 1) problem formulation, which builds a
model for each studied system to represent the relationship between its
object popularities, resource constraints and targeted performance metrics
as a constrained optimization problem; 2) optimization, which derives the
optimal or approximate solution to the problem by using techniques like
Lagrange multipliers, dynamic programming, linear programming relaxation,
and random walks guided by the Metropolis-Hastings algorithm; 3) analytical
results on the complexity of the optimization problem, the asymptotic
optimization time, and the approximation ratios of approximate solutions.

On the systems side, we explore the effectiveness and feasibility of
popularity-driven data management on running traces collected from real,
large systems. We also address a number of issues that are key to the
implementation of popularity-driven strategies in practice, e.g., dynamic
object popularity changes, offline computation cost, etc.
</description>
    <pubDate>Tue, 26 Jun 2007 11:09:12 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/209</guid>
   </item>
  
   <item>
    <title>Chengliang Zhang: &quot;Behavior Models for Computer Programs&quot;</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/210</link>
    <description>[Tuesday, July 17, 2007 at  2:00 PM in Computer Science Bldg. Room 601] Today&#39;s software systems are becoming more and more complex. In
  many senses, a software system is very similar to a live organism. 
  This motivates us to study a software system by investigating its
  behavior. A program&#39;s behavior is decided by not only the code and
  data defined by the program itself, but also the user inputs and
  the execution environment such as the hardware, the operating system, 
  and possibly multiple layers of virtual machines.

  Program behavior analysis complements the traditional code-based
  analysis such as data and control flow analysis and dependence 
  theory. Deeper understanding of the program&#39;s behavior can help us
  improve performance by better utilizing the modern hardware.
  Based on the traces of a program execution, we build statistical
  models and identify recurring large-scale patterns. These models are 
  then used to predict the future program behavior and expose
  opportunities for program optimization.

  In this thesis, we investigate three behavior models: reference
  affinity, data locality quality, and program phases. We prove 
  the hardness of reference affinity and show its application in
  divide-and-conquer type computations. We also propose a
  sampling-based method to  find approximately the reference affinity
  groups. We propose a new component-based definition of spatial 
  locality. With the locality we attempt to find the limits to
  improvement of spacial locality.

  Based on the behavior models, we adapt programs to improve their
  performance. From the spatial aspect, we propose to improve the 
  layout of matrix multiplication and code layout based on
  reference affinity. From the resource aspect, we propose a new
  resource-based memory management scheme and its extension in
  multi-programming environments. From the temporal aspect, we propose 
  to monitor program memory usage, detect memory leaks and trigger preventive
  garbage collection based on the program phase behavior. We also
  propose layout auditing to improve the spatial data locality by
  continuously measuring and selecting the best garbage collection
  method during a long-running Java program execution.
</description>
    <pubDate>Wed, 13 Jun 2007 15:03:14 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/210</guid>
   </item>
  
   <item>
    <title>Ding Liu: &quot;Statistical Semantic based Machine Translation&quot;</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/207</link>
    <description>[Friday, May 11, 2007 at  2:00 PM in Computer Science Bldg. Room 601] Statistical Syntax based Machine Translation (MT) has been well developed and achieved competitive
performance as the state-of-art phrase based MT systems. It is due to our better understanding of the
synchronous
grammars and better techniques on searching and pruning. As we all know, semantics is a more general structure
built upon the syntax, and the semantic roles show the functionality of the syntactic constituents with
respect to a certain predicate.
With the techniques for the syntax based MT, we cannot help but wonder: can we go one step further and make a
statistical semantic based MT sytem?

In this proposal, we address the potential benefit of using semantic features in a MT system, and the possible
ways of building a statistical semantic based MT system upon a syntax based MT system. The tree-to-string
template approach to the syntax based MT
provides a good framework to incorporate the semantic information, with which our semantic based MT can be
divided
into two steps: first, extract the semantic role information from the syntax trees, second,
use the semantic features to modify the syntax MT models. We propose several methods to integrate the semantic
role
features into a tree-to-string MT system, such as enriching the tree node labels with the semantic role labels,
constructing versatile semantic features for the discriminative MT models, and making a two-layer translation
model with the semantic structure.

We also explore
the new methods for the manual/automatic evaluation of semantic based MT. The motivation is that the semantic
features can be used to divide the evaluation task into two parts: the skeleton evaluation focusing on fluency,
and the arguments evaluation focusing on adequacy. Such division makes the manual evaluation more managable and
provides the automatic evaluation features in semantic level.
</description>
    <pubDate>Wed, 09 May 2007 12:17:54 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/207</guid>
   </item>
  
   <item>
    <title>Pin Lu: &quot;Hypervisor Support of Data Caching and Crash in Virtual Machine Architecture&quot;</title>
    <link>http://www.cs.rochester.edu/dept/seminar/view/206</link>
    <description>[Monday, May 14, 2007 at  3:30 PM in Computer Science Bldg. Room 601] Virtual Machine (VM) technologies have recently started to experience a resurgence in both research and industry communities. Many desirable features of the VM, including fault containment, performance isolation, resource consolidation, ease of management and migration, and checkpointing, are of the interests of various applications nowadays.
This thesis work will be focusing on hypervisor-level cache design and utilization, and hypervisor support for data recovery upon operating system (OS) crashes.

In this proposal, we first investigate a hypervisor-level data caching approach. In this approach, the hypervisor takes over the management for part of the VM memory.  It manages the memory as an exclusive cache to avoid double caching. The cache admits data directly from the VM memory to save overhead. We implemented and evaluated the proposed hypervisor exclusive cache. Our current design requires guest OS changes and the prototype implementation is not fully optimized. The future work is to improve the transparency and performance of the current scheme.

Our hypervisor exclusive cache provides a means to trace VM data access pattern. We describe a utilization of the hypervisor exclusive cache in obtaining more complete page miss ratio curve (MRC) of a VM, which in turns guides VM memory allocation with flexible performance isolation constraints. The evaluation results are presented in comparison with VMware ESX server. We plan to further investigate other uses of the hypervisor exclusive cache including hypervisor-level caching and prefetching.

Finally, we illustrate our preliminary thoughts on hypervisor support for
data recovery.  When employing delayed writes, an OS crash may lose buffered data for delayed writes, resulting in a loss of data durability. Despite the attempt to support data recovery at the OS level, we argue that moving the recovery mechanism out of the OS has certain architectural advantages. Possible challenges are discussed and investigation plan is presented.
</description>
    <pubDate>Wed, 09 May 2007 12:14:52 -0400</pubDate>
    <guid>http://www.cs.rochester.edu/dept/seminar/view/206</guid>
   </item>
  
 </channel>
</rss>
