program systems institute russian academy of sciences1 program systems institute research activities...
TRANSCRIPT
Program Systems Institute Russian Academy of Sciences
11
Program Systems InstituteProgram Systems InstituteResearch Activities OverviewResearch Activities Overview
Extended Version Extended Version
Program Systems InstituteProgram Systems InstituteResearch Activities OverviewResearch Activities Overview
Extended Version Extended Version
Alexander Moskovsky,Alexander Moskovsky,Program Systems Institute RAS, Program Systems Institute RAS,
Russia, Pereslavl-ZalesskyRussia, Pereslavl-Zalessky
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
22
Institute Institute StructureStructure Institute Institute StructureStructure ~200 research staff~200 research staff Artificial Intelligence Research CentreArtificial Intelligence Research Centre Medical Informatics Research CentreMedical Informatics Research Centre Research Center for Research Center for
Multiprocessor SystemsMultiprocessor Systems System Analysis Research CentreSystem Analysis Research Centre Control Processes Research CentreControl Processes Research Centre Scientific and educational centreScientific and educational centre::
International Children’s Computer International Children’s Computer Centre named after A. AylamazyanCentre named after A. Aylamazyan
Kindergarten and Primary school Kindergarten and Primary school “Pochemuchka”“Pochemuchka”
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
““SKIF-GRID” PROJECT TIMELINESKIF-GRID” PROJECT TIMELINE““SKIF-GRID” PROJECT TIMELINESKIF-GRID” PROJECT TIMELINE
PSI RAS PSI RAS is a lead organizarion in Russian Federationis a lead organizarion in Russian Federation
1.1. 2000-2004 - 2000-2004 - SKIF project, SKIF K-1000 is #98 in SKIF project, SKIF K-1000 is #98 in Top500Top500
2.2. JuneJune 2004 2004 – – first proposal filedfirst proposal filed for “SKIF-GRID” for “SKIF-GRID” projectproject
3.3. MarchMarch 2007 2007 – – approved by Governmentapproved by Government
4.4. MarchMarch 200 20088 - - SKIF-MSU supercomputer deployed SKIF-MSU supercomputer deployed (#36 in June 08 Top 500)(#36 in June 08 Top 500)
5.5. May 2008May 2008 - “SKIF-Testbed” federation created. - “SKIF-Testbed” federation created.
6.6. March 2009March 2009 – alliance agreement signed for SKIF – alliance agreement signed for SKIF series 4 developmentseries 4 development
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
SKIF MSU SKIF MSU SKIF MSU SKIF MSU
Theoretical peak Theoretical peak performance performance 60 TFlops60 TFlops
47 TFlops 47 TFlops Linpack Linpack Advanced clustering Advanced clustering
solutionssolutions:: diskless diskless
computational computational nodesnodes
Original blade Original blade designdesign
Parameter Value
CPU architecture: x86-64
CPU model: Intel XEON E5472 3,0 GHz (4-cores)
Nodes (dual CPU) 625
CPU cores total 5 000
Interconnect Infiniband DDR,
Fat Tree
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
PROJECT ORGANIZATION: 2009-PROJECT ORGANIZATION: 2009-20102010
PROJECT ORGANIZATION: 2009-PROJECT ORGANIZATION: 2009-20102010
Project directionsProject directions1.1. Grid technologyGrid technology
2.2. SupercomputersSupercomputers SWSW HWHW
3.3. SecuritySecurity
4.4. Pilot projects – applications of Pilot projects – applications of HPC and grid technologyHPC and grid technology
TotallyTotally more than 20 more than 20 organizations in Russiaorganizations in Russia
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
SKIF-Aurora SKIF-Aurora SKIF-Aurora SKIF-Aurora Fulfillment of original SKIF-GRID project Fulfillment of original SKIF-GRID project
goals (back to 2004)goals (back to 2004) Highest density of performanceHighest density of performance
(increase number of FLOPS per 1U)(increase number of FLOPS per 1U) Interconnect:Interconnect: we need better scalability, we need better scalability,
bandwidth and latency that it’s provided by bandwidth and latency that it’s provided by best available solutions (eg. Infiniband QDR)best available solutions (eg. Infiniband QDR)
New approach to monitoring and New approach to monitoring and management management of the supercomputerof the supercomputer
CPUs and acceleratorsCPUs and accelerators in computational in computational nodes of the supercomputernodes of the supercomputer
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
SKIF Aurora OverviewSKIF Aurora OverviewSKIF Aurora OverviewSKIF Aurora Overview Nodes: Nodes:
dual-CPU boards (Nehalem)dual-CPU boards (Nehalem) FPGA for 3D torus interconnectFPGA for 3D torus interconnect Infiniband QDR integratedInfiniband QDR integrated liquid coolingliquid cooling
Chassis: Chassis: 32 nodes32 nodes Integrated management and Integrated management and
monitoring monitoring Integrated Integrated Infiniband QDRInfiniband QDR Cables integrated on backplaneCables integrated on backplane
Rack:Rack: Up to 8 chassis (512 CPU)Up to 8 chassis (512 CPU) No moving parts No moving parts
Designed by an alliance of Eurotech, PSI RAS Designed by an alliance of Eurotech, PSI RAS and RSC SKIF with support by Inteland RSC SKIF with support by Intel
chassis
RackDatacenter
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
SKIF-4 nodeSKIF-4 nodeSKIF-4 nodeSKIF-4 node
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
Subsidiary Interconnect, Infiniband
Our proposalOur proposalOur proposalOur proposal
Based on SKIF-Aurora architecture:Based on SKIF-Aurora architecture: Interconnect w. 3-D torus topologyInterconnect w. 3-D torus topology FPGA on network levelFPGA on network level
99
System Interconnect, 3D-torus
FPGA FPGA FPGA FPGA...
CPU CPU CPU CPUstandard part
non-standard part
Program Systems Institute, Russian academy of ScienceProgram Systems Institute, Russian academy of Science
Our proposal-2Our proposal-2Our proposal-2Our proposal-2 FPGA-basedFPGA-based Integral support of Integral support of
programming programming model by hardwaremodel by hardware
Customizable hw Customizable hw platformplatform
Project idea: Project idea: optimized runtime optimized runtime support for FP-support for FP-based languagesbased languages
1010