Colloquia and SeminarsTo join the email distribution list of the cs colloquia, please visit the list subscription page.
- Bioinformatics Forum
- BizTEC Forum
- CGGC Weekly Seminar
- Haifux, Haifa Linux Club
- Pixel Club
- Theory Seminar
Upcoming Colloquia & Seminars
ceClub: Catapulting beyond Moore’s Law: Using FPGAs to Accelerate Data Centers
- Derek Chiou (Microsoft)
- Monday, 27.4.2015, 11:30
- EE Meyer Building 861
In this talk, I will describe a joint Microsoft Research and Bing project to study the possibility of using field programmable gate arrays to accelerate cloud applications. We developed an FPGA card that plugs into a Microsoft designed cloud server and used it to accelerate a significant portion of Bing’s search engine. Because the application does not fit into a single FPGA, we spread the application across multiple FPGAs spread across multiple servers and connected together with a low latency network also implemented by the FPGA. Our 1,632 server pilot demonstrated a factor of two performance improvement over pure software at significant power savings. FPGAs will accelerate Bing searches in one data center starting in 2015.
Derek Chiou is a Principal Architect at Microsoft where he co-leads a team working on FPGAs for data center applications and an associate professor at The University of Texas at Austin. His research areas are FPGA acceleration, high performance computer simulation, rapid system design, computer architecture, parallel computing, Internet router architecture, and network processors. Before going to UT, Dr. Chiou was a system architect at Avici Systems, a manufacturer of terabit core routers. Dr. Chiou received his Ph.D., S.M. and S.B. degrees in Electrical Engineering and Computer Science from MIT.
Pixel Club: Interactive Annotation by Randomly Perturbing Max-solvers
- Tamir Hazan (CS, Haifa University)
- Tuesday, 28.4.2015, 11:30
- EE Meyer Building 1061
Predictions in modern inference problems can be increasingly understood in terms of high-dimensional structures such as arrangements of objects in computer vision, parses in natural language processing or molecular structures in computational biology. To interactively annotate an image, one needs to efficiently sample possible interactions to suggest the user to annotate areas with high uncertainty. Unfortunately, traditional approaches to handle high-dimensional structures involve the Gibbs distribution for which sampling and inference are notoriously hard. In this talk I will present recent advances in machine learning that facilitate sampling and probabilistic reasoning in high-dimensional problems. This approach constructs new distributions over high-dimensional structures that randomly perturb the image information followed by applying graph-cuts (max-solver).
ceClub: Scaling Concurrent Log-Structured Data Stores
- Eshcar Hillel (Yahoo! Labs)
- Wednesday, 29.4.2015, 11:30
- Taub 9
Log-structured data stores (LSM-DSs) are widely accepted as the state-of-the-art implementation of key-value stores. They replace random disk writes with sequential I/O, by accumulating large batches of updates in an in-memory data structure and merging it with the on-disk store in the background. While LSM-DS implementations proved to be highly successful at masking the I/O bottleneck, scaling them up on multicore CPUs remains a challenge. This is nontrivial due to their often rich APIs, as well as the need to coordinate the RAM access with the background I/O.
We present cLSM, an algorithm for scalable concurrency in LSM-DS, which exploits multiprocessor-friendly data structures and non-blocking synchronization. cLSM supports a rich API, including consistent snapshot scans and general non-blocking read-modify-write operations. We implement cLSM based on the popular LevelDB key value store, and evaluate it using intensive synthetic workloads as well as ones from production web-serving applications. Our algorithm outperforms state of the art LSMDS implementations, improving throughput by 1.5x to 2.5x. Moreover, cLSM demonstrates superior scalability with the number of cores (successfully exploiting twice as many cores as the competition).
This is a joint work with Guy Golan, Eddie Bortnikov (Yahoo Labs) and Idit Keidar (Technion, Yahoo Labs), presented at EuroSys 2015.
Eshcar Hillel is a Research Scientist at Yahoo Labs, currently working in the Scalable Search Systems Research team. Eshcar received the B.Sc. degree in Software Engineering and the Ph.D. degree in Computer Science from the Technion, in 2002 and 2011, respectively. Prior to her graduate studies she worked at Hewlett-Packard as part of the Non-Stop relational database system team.
I/O Page Faults
- Ilya Lesokhin, M.Sc. Thesis Seminar
- Wednesday, 29.4.2015, 12:30
- Taub 601
- Prof. D. Tsafrir
Most modern computer systems use demand paging based virtual memory. Address translation for virtual memory support in I/O devices is known to be useful. It allows SR-IOV (single root I/O virtualization) direct device assignment in virtualization scenarios and kernel bypassing access to I/O devices by user space applications. However, these mechanisms require memory pinning and strict, static memory allocation as they do not support demand paging. This deficiency prevents memory overcommitment, complicates the programming model and hinders administration. This paper presents a working hardware and software support for I/O page faults (IOPFs) and demand paging in a network interface card (NIC). Demand paging by the I/O devices imposes unique challenges, which differ significantly from the demand paging on the CPU. After tackling these challenges, our IOPF enabled NIC allows for efficient memory over-commitment. In a number of real world benchmarks, the memory over-commitment improve performance by 1.2x--1.9x compared to the same systems using memory pinning.
Theory Seminar: Cops, Robbers, and Threatening Skeletons (Padded Decomposition for Minor-Free Graphs)
- Ofer Neiman (Ben-Gurion University)
- Wednesday, 29.4.2015, 12:30
- Taub 401
We prove that any graph excluding $K_r$ as a minor can be partitioned into clusters of diameter at most $\Delta$ while removing at most $O(r/\Delta)$ fraction of the edges. This improves over the results of Fakcharoenphol and Talwar, who building on the work of Klein, Plotkin and Rao gave a partitioning that required to remove $O(r^2/\Delta)$ fraction of the edges.
Our result is obtained by a new approach to relate the topological properties (excluding a minor) of a graph to its geometric properties (the induced shortest path metric). Specifically, we show that techniques used by Andreae in his investigation of the cops-and-robbers game on excluded-minor graphs can be used to construct padded decompositions of the metrics induced by such graphs. In particular, we get probabilistic partitions with padding parameter $O(r)$ and strong-diameter partitions with padding parameter $O(r^2)$ for $K_r$-free graphs, padding $O(k)$ for graphs with treewidth $k$, and padding $O(\log g)$ for graphs with genus $g$.
Joint work with: Ittai Abraham, Cyril Gavoille, Anupam Gupta and Kunal Talwar.
Startup Fair and Recruitment at CS
- Wednesday, 29.4.2015, 12:30
- CS Taub Lobby & Auditorium 2
CS invites you to a STARTUP FAIR and recruitment by the presenting firms: Click-tale, Forter, Kamimnario, Ravello, Outbrain, Taxi, Wix.com and more. In addition, lectures will be given by the firms representatives.
The event will take place on Wednesday, April 29, between 12:30-14:30 at the CS Taub Lobby. The lectures will be held in Auditorium 2.
You are all welcome!
Amdocs Best Project Contest - The Finals
- Wednesday, 29.4.2015, 16:30
- CS Taub Auditorium 2
You are invited to the final stage of the 2014 Amdocs Best Project Contest. The competing teams will present and talk about their projects.
The event will take place on Thursday, April 29, 2015, 16:30-18:30, in Auditorium 2, CS Taub Building.
You are all invited to cheer, support and be exposed to outstanding projects.
Coding for New Applications in Storage Media
- Artyom Sharov, Ph.D. Thesis Seminar
- Wednesday, 29.4.2015, 16:30
- Taub 601
- Prof. Ronny Roth
Conventional magnetic recording media are composed of basic magnetizable two-dimensional units called grains, which might be random in size and shape. Recently, a new technological enhancement was proposed, which enables magnetizing areas as small as the size of grains. This novelty effectively created a different type of medium, in which one observes a new type of errors. Handling such errors is the main subject of this work. We first consider a combinatorial model of this medium, where so-called grain-correcting codes are used to handle worst-case error patterns. We also study a variant of this model where the errors are allowed to overlap (the supporting application of this variant can be found in shingled writing on bit-patterned media). For these two models, we present improved lower and upper bounds on the size and the growth rate of grain-correcting codes. In addition, we give explicit constructions of grain-correcting codes for correcting very small or very large number of errors. We then switch to the problem of detecting grain errors and provide lower and upper bounds on the minimum redundancy of codes that can detect any number of grain errors. Then, we turn to a probabilistic characterization of overlapping grain error patterns as a generalized Ising channel, for which we give almost-tight bounds on the capacity. We also consider a variant of this channel where feedback is added and compute lower bounds on its capacity. Moreover, for a certain range of values of the channel error probability, we establish a closed-form expression for that capacity.
CGGC Seminar: Solid Sweeps in Computer Aided Geometric Design
- Jinesh Machchhar (CS, Technion)
- Sunday, 3.5.2015, 13:30
- Room 337-8 Taub Bld.
Solid sweep is the envelope of the swept volume of a given solid moving along a one parameter family of rigid motions in 3 dimensional space. Solid sweep is a powerful and versatile surface design tool which has found uses in many areas like CNC-machining verification, collision detection and packaging. This work aims to develop a robust computational framework for sweeps in the realm of boundary representation (brep) format, wherein, the input solid as well as the outputted swept volume are represented in brep. This involves the following considerations: (i) Giving an accurate parametrization of the faces and edges in the brep of the swept volume, (ii) Resolving self-intersections and singularities, and (iii) Computing topological information, viz., adjacency relations amongst the faces, edges and vertices of the swept volume as well as the orientation of these.
In almost all non-trivial sweeps, there exists no closed form parametrization of the envelope surface. We address this problem via the procedural paradigm in which the parametric definitions of curves and surfaces are stored as numerical procedures.
We give a novel classification of sweeps into simple, decomposable and non-decomposable based on the complexity of trim regions. A novel geometric invariant aids in efficient classification as well as in locating the trim curves.
We show that while the global brep structure of the input solid and that of the envelope are quite different, locally they are similar. This aids in lifting the brep structure of the solid to that of the envelope.
(Based on joint work with Milind Sohoni and Bharat Adsul).
Oblivious Algorithms - How to Serve in the Dark
- Yossi Azar - Colloquium Lecture
- Tuesday, 5.5.2015, 14:30
- Room 337-8 Taub Bld.
First Project Fair at CS
- Tuesday, 26.5.2015, 15:30
- CS Taub Building
We are happy to invite you to the first project fair at CS on Tuesday, May 26 2015, between 15:00-18:00 at the CS Taub lobby,
The event will host top executives from the hi-tech industry. The participating students will get a great opportunity to present their projects accomplished throughout their studies. Furthermore, the firms representatives will be exposed to the applicable aspects of the degree. and to the students' development and implementing abilities, and thus to foresee their potential in future cooperation.
Participating is free but requires pre-registraion.
You are all invited!
CS M.Sc. & Ph.D. Alumni Event
- Thursday, 11.6.2015, 15:00
- CS Taub
We are happy to invite you to all times MSc & PhD CS alumni event.
The event will be held on Thursday, June 11th at the CS Taub Building, starting at 15:30 and will include a fascinating workshop with famous speakers: Prof. Amnon Shashua, Prof. Hossam Haic, Prof. Michael Aviram and Dr. Haim Shapira and from 19:00, a social and networking happening with guest artist Liior Schuard. program and pre-registration.
If you are familiar with other MSc & PhD alumni, we would be happy to receive from you their contact details to ensure ultimate participants.
Hope to see you!br>