Date post: | 19-Jun-2015 |
Category: |
Documents |
Upload: | cesga-foundation |
View: | 198 times |
Download: | 0 times |
SVG Upgrade 2012
AGENDA
•Upgraded SVG
• Motivation
• Hardware configuration
• Infiniband Network
• Environment
• Queues configuration
• Benchmarks
•CESGA Supercomputers in 2013
• Distribution of jobs
Hardware
Thin-nodes 8 HP SL230s Gen8 each with
2x Intel Xeon E5-2670, 8 cores each, 2.6GHz
64GB main memory DDR3-1600MHz
2TB SAS hard disk
2x1GbE
2xInfiniband FDR 56Gb
Peak Performance 332GFlops
Fat-nodes 2 HP DL560 Gen8 each with
4x Intel Xeon E5-4620, 8 cores each, 2.2GHz
512GB main memory DDR3-1600MHz
6 hard disks each 1TB
4x1GbE
Infiniband FDR 56Gb
10GbE
Peak Performance 563 GFlops
Total 24 Intel Sandy Bridge processors 192 cores 1,5 TB memory 28 TB disk Peak Performance 3788 GFLops
Total
24 Intel Sandy Bridge CPUs
192 cores
1,5 TB memory
28 TB disk
3788 Gflops peak performance
Motivation
Target
• Competitive solution for:
• Parallel MPI & OpenMP applications
• Memory Intensive
• Alternative for Finis Terrae
• Lower cost of operation and maintenance
• Finis Terrae II prototype
• To define new requirements
Infiniband network
Mellanox SX6036 switch
36 port FDR 56Gb/s
4Tb/s aggregated non-blocking BW
1 microseconds MPI latency
Dual connection:
High availability – same BW
Environment
Integrated in the SVG cluster:
Scientific Linux 6.3 (Red Hat)
Common /opt/cesga
Common /home, stores…
Same gateway: svg.cesga.es
Interactive use: compute –arch sandy
Jobs: qsub –arch sandy
Binary compatible - no need to recompile
Usage
svg.cesga.es
SSH
compute –arch amd qsub –arch amd
compute –arch sandy qsub –arch sandy
compute qsub
Configuration
Full production phase (November 2012)
• Only runs jobs with –sandy option
• General availability of applications specifically compiled
• Maximum wall-clock time 12 hours
• Maximum 2 jobs per node fat nodes
Under consideration near future:
Jobs without –sandy option
Higher wall-clock time
Queues Configuration
Exclusive nodes
To take advantage of Infiniband
To take advantage of Turboboost
Jobs not interferring each other
Maximum performance
Maximum 2 jobs on Fat nodes
32 cores nodes
Exclusive if required by the jobs (cores, memory)
Queues: Limits “module help sge”
Up to 112 cores (MPI)
Up to 32 cores shared memory (OpenMP)
Memory:
up to 64GB per core
up to 512GB for non MPI jobs
up to 1024GB per job
Scratch up to 1,7TB
Execution time: 12 hours
If needed ask for more resources in https://www.altausuarios.cesga.es/
Intel® Xeon® Processor E5-2600 Turbo Boost
1Max Turbo Boost frequency based on number of 100 MHz increments above marked frequency (+1 = 0.100 GHz, +2 = 0.200 GHz, +3 = 0.300 GHz, etc. )
CESGA Supercomputers 2013
Finis Terrae (FT)
Capability computing
Parallel jobs (>4 ... 1024 cores)
Huge memory (>4... 1024GB)
Huge parallel scratch (>50... 10,000GB)
Superordenador Virtual Gallego (SVG)
Throughput and Capacity computing
Sequential & parallel jobs up to 32 cores per node and 112 cores
MPI
Low-medium-large memory (up to 512GB!)
Medium single node scratch (<1000GB)
Customized clusters – Cloud services
Shared storage: /home
/store
Linux O. S.
Grid Engine Batch Scheduler
Other Improvements
VPN for home connection
Storage:
Do not use SFS from SVG
Use “store”
High availability front-ends:
svg.cesga.es