Seattle Conference on Scalability: Building a Scalable Resource Management
Google engEDU
56 min – Jun 23, 2007

Google Tech Talks
June 23, 2007

ABSTRACT

2007 Google Seattle Conference on Scalability: Building a Scalable Resource Mgmt System for Grid Computing
Speaker: Khalid Ahmed, Platform Computing Corp.

This talk will describe the architecture and implementation details
for building a highly scalable resource management layer that can
support a variety of applications and workloads. This technology
has evolved from large scale computing grids deployed in
production at customers such as Texas Instruments, AMD, JP
Morgan, and various government labs. We will show how to build a
centralized dynamic load information collection service that can
handle up to 5000 nodes/20,000 cpus in a single cluster. The
service is able to gather a variety of system level metrics and is
extensible to collect up to 256 dynamic or static attributes of a node
and actively feed them to a centralized master. A built-in election
algorithm ensures timely failover of the master service ensuring
high-availability without the need for specialized interconnects.
This building block is extended to multiple clusters that can be
organized hierarchically to support a single resource management
domain that can span multiple data centers. We believe the current
architecture could scale to 100,000 nodes/400,000 cpus. Additional
services such as a distributed process execution service, and a
policy-based resource allocation engine which leverage this core
scale-out clustering service are described. The protocols,
communication overheads, and various design tradeoffs that were
made the development of these services will be presented along
with experimental results from various tests, simulations and
production environments.
Khalid Ahmed works as the Chief Architect and Director of
Technology Research at Platform Computing. In over 12 years at
Platform he worked in a number of roles including development,
product management and architecture. His work on distributed
scheduling, wide-area resource sharing, workload management,
system automation, virtualization management, and high availability Read the rest of this entry »

Motion Estimation from Image and Inertial Measurements
Google engEDU
56 min – Feb 24, 2006

Google TechTalks
February 24, 2006

Dennis Strelow

Robust motion from monocular image measurements would be an enabling technology for Mars rover, micro air vehicle, and search and rescue robot navigation; and for modeling complex environments from video.
While algorithms exist for estimating six degree of freedom motion from monocular image measurements, motion from image measurements suffers from inherent problems. These include sensitivity to incorrect or insufficient image feature tracking; sensitivity to camera modeling and calibration errors; and long-term drift in scenarios with missing observations, i.e., where image features enter and leave the field of view.
The integration of image and inertial measurements is an attractive solution to some of these problems. Among other advantages, adding inertial measurements to image-based motion estimation can reduce the sensitivity to incorrect image feature tracking and camera modeling errors. On the other hand, image measurements can be exploited to reduce the drift that results from integrating noisy inertial measurements; and allow the additional unknowns needed to interpret inertial measurements, such as the gravity direction and magnitude, to be estimated. Read the rest of this entry »

PhotoSketch: A Photo-Centric 3D Modeling Tool
Google engEDU
56 min – Jun 26, 2007

Google Tech Talks
June 26, 2007

ABSTRACT

Geospatial navigation tools such as Google Earth and Microsoft Virtual Earth are exceedingly popular applications for exploring massive datasets. Their explosive growth provides impetus for photorealistic 3D modeling of urban scenes. Although laser range scanners are traditional sources for detailed 3D models of existing structures, the heavyweight models they generate are not appropriate for the streaming data that these navigation applications leverage. Instead, lightweight models as produced by photogrammetry tools are better suited for this domain. Unfortunately, photogrammetry tends to be a very labor-intensive operation, requiring manual camera calibration and tedious modeling processes. This talk presents the virtues of combining computer vision techniques and photogrammetry to simplify the 3D modeling workflow. The contribution of this work is that it merges the benefits of automatic feature extraction, multiview geometry, an intuitive sketching interface, and dynamic texture mapping to produce lightweight photorealistic 3D models of buildings. We present results from experiments in large-scale urban scenes.

Speaker: George Wolberg Read the rest of this entry »

Navigating the World’s Photographs
Google engEDU
1 hr 1 min – Jun 13, 2007

Google Tech Talks
June 13, 2007

ABSTRACT

There’s a big difference between looking at a photograph of a place and being there. But what if you had access to every photo ever captured of that place and could conjure up any view at will? With billions of photographs currently available online, the Internet is beginning to resemble such a database, capturing most of the world’s significant sites from a huge number of vantage points and viewing conditions. For example, a Google image search for "notre dame" or "grand canyon" each returns more than half a million photos, showing the sites from myriad viewpoints, different times of day and night, and changes in season, weather and decade.

This talk explores ways of transforming this massive, unorganized photo collection into visualizations of the world’s sites, cities, and landscapes. After a brief recap of our work on Photo Tourism and Photosynth, I will focus on current efforts and newest results, in the domains of 3D scene reconstruction and new visual interfaces for navigating photo collections. Read the rest of this entry »

The Theoretical Limits of Statistical High Dimensional Nearest Neighbor Algor…

55 min – Sep 6, 2007

Google Tech Talks
September 6, 2007

ABSTRACT

Suppose we have $n$ Bernoulli(1/2) long sequences of bits. Let $n-2m$ sequences be completely independent, while the remaining $2m$ sequences are composed of $m$ independent pairs. The interdependence within each pair is that their bits agree with probability $1/20$. The exponent $1/p$ is optimal in a large natural class of algorithms which we name Bucketing Codes. Moreover if one sequence out of each pair belongs to a known set of $n^{(2p-1)^{2}-epsilon}$ sequences, than pairing can be done using order $n$ comparisons!

These results are extended to a general discrete independent data model. The performance of Bucketing Codes is bounded by a newly defined Bucketing Information. That bound can be asymptotically attained by randomly constructed bucketing codes. Our Information Theory inspired approach is contrasted with the worst case analysis dominating the literature.

Speaker: Moshe Dubiner Read the rest of this entry »