1. RECONFIGURABLE DATA PROCESSING FOR CLOUDS Parth Shah Kanika Chawla Sowmith Boyanpalli 1

2. MOTIVATION To Implement Reconfigurable computing in Cloud Based Systems 2 / 20

3. INTRODUCTION  Cloud Computing??  Datacenters??  Reconfigurable Computing?? 3 / 20

4. BACKGROUND  Reconfigurable computing paved way for mainstream high performance computing.  Modern FPGA’s implement high bandwidth custom memory hierarchies offering flexibility and performance  Implementing these FPGA’s in modern computing systems is hindered by the need for specialized FPGA boards and also lack of programming modes  They also don’t have a good O.S suport and ABI interfaces for application development. 4 / 20

5. WHAT’S IN THE PAPER  Using FPGA’s just like CPU’s and GPU’s in the cloud.  Using FPGA’s in datacenters scaling down the energy requirement when compared to using conventional processors. 5 / 20

6. CLOUD COMPUTING  History  Virtualization  Xen Hypervisor – Partition Multiple Guest OS – adopted by amazon  Sudden surges in load is handled by adding additional virtual machines -Providing dynamic reconfiguration 6 / 20

7. HARDWARE CLOUDS  Data Centers employ horizontal scaling by increasing the number of hosts.  Vertical Scaling is difficult as modern systems are made up of multicore systems with constant clock speed.  The growth of data centers is limited due to energy constraints and software has reached the saturation.  Hence Hardware improvisation is needed. 7 / 20

8. RESEARCH INTERESTS  Major research interests areas in cloud computing  Impact of reconfigurable FPGA on these research areas 1. Operating Systems 2. Datacenter Programming 3. Information Security

9. OPERATING SYSTEMS  Hypervisors expose simple network and storage interfaces to VMs.  Actual Physical drivers handled elsewhere 9 / 20

10. OPERATING SYSTEMS  Networking involves bridging, topology and integration with systems like Openflow  Storage deals with Snapshots and deduplication of common blocks across VMs  Currently limited to I/O, can be extended to run computation over data  Availability of GPU and Programmable I/O boards  New OS designed to simplify the task of writing, deploying and profiling code across heterogeneous platforms 10 / 20

11. DATACENTER PROGRAMMING  Distributed dataflow frameworks that transparently handle fault tolerance, resource scheduling, synchronization, message passing  Examples: - MapReduce, Dryad, CIEL  Simple Programming models  Directed Acyclic Graphs (DAG)  Run-time schedules iteratively walks DAG  Prepares host to ensure data is available locally  Main challenge: Reconfiguring FPGAs  Mesos investigates partitioning physical resources across multiple frameworks on same set of hosts 11 / 20

12. DATACENTER PROGRAMMING  Encouragement for novel database models other than SQL and ACID models  Experiments done to program all data processing operators on top of large FPGAs  Right computation model enables significant improvement in power consumption and parallelization  Close integration between high level host languages and FPGAs required  MORA: an example of DSL for streaming vector and matrix operations aimed at multimedia applications.  Compiler infrastructures like LLVM to convert C/C++ code to FPGAs  Less effort to design and implement portably 12 / 20

13. DATACENTER PROGRAMMING  Many tools and methods available for easy programming and synthesis  Accelerator: library to synthesis data parallel programs in C# to FPGAs  Multi-stage programming: Abstract algorithms converted to high level language and then into desired architecture. 13 / 20

14. INFORMATION SECURITY

15. 15/20  CLOUD outsources processing over large datasheets  This code is often written in C or FORTRAN  Even a small bug in input can let attackers execute arbitrary code on the host machine  Although the attackers cannot access the virtual data in the hypervisor, they still have access to many local network resources  Malicious data can be crafted to exploit memory errors and execute as a code on a CPU THE ISSUE

16. THE SOLUTION 16 / 20  By shifting from software to hardware, information security is improved immensely.  Baking algorithm implementation into FPGA ensures that attackers cannot run their arbitrary code in the system.  The data is never manipulated via the CPU rather it is directly compiled to the FPGA  Only a small channel exists between OS and FPGA to communicate results.  The conventional FPGA model has scope of hacking through SRAM but when deployed in cloud, attackers don’t have any physical access.

17. DATA SECURITY  Data needs security against the untrusted cloud infrastructure, which can be ensured using encoding data.  Encoding is done for information flow constraints, which are ideal for domain specific data flow languages.  Homomorphic encryption are also being deployed. It permits computation over encrypted data without decrypting the underlying data.  But it is expensive to implement.  But recently, lattice based cryptography which decrease the complexity cost of homomorphic encryption, are being used.  Using Lattice reduction, speedups of 2.12 have been reported. 17 / 20

18. DATA SECURITY  Reducing cost of cryptography has a social impact.  Until now anonymity networks such as Tor AND Freenet were being used for storing data. Using them is way too slow.  There have been talks to shift this burden to cloud computing, but cost remains the main constraint.  Virtual networking is reconfigured much more than hardware setups. 18 / 20

19. THE FUTURE CHALLENGES  How to unify the demands of data-centric processing, language integration, network processing into a single infrastructure  The need for better OS integration, device models, and abstractions  Without an ABI or source API, software re-use and integration is very difficult.  Debugging and visualization support. General purpose systems provide a hypervisor-kernel-user space language runtime model that gets progressively easier and higher-level to debug. Abstraction boundaries exist where they don’t in current FPGAs. Staged programming or functional testing in a general-purpose systems makes this easier.  We need to develop a common set of concepts, principles and models for application execution on reconfigurable computing platforms to allow collaboration between universities and companies and to provide a solid framework to build new innovations and applications. This kind of eco-system has been sadly lacking for reconfigurable systems. 19

20. CONCLUSION  It is encouraging that cloud computing is driven by fine-grained charging for the computation resources used.  Reconfigurable FPGAs driven down the cost of many types of computation commonly found on the cloud, and thus a community-driven deployment of a cloud setup with rentable hardware would provide a focal point to “fill in the blanks” for reconfigurable FPGA computing in the cloud. 20 / 20

21. LIMITATIONS  The paper does not provide any experimental results and is much of a literature survey  Half of the paper covers background and introduction and does not give much information about implementation.  A practical model or example should have been given.

22. THANK YOU