High Performance Computing G Burton – ICG – Oct12 – v1.1 1.
-
Upload
pearl-phelps -
Category
Documents
-
view
216 -
download
0
Transcript of High Performance Computing G Burton – ICG – Oct12 – v1.1 1.
High Performance Computing
G Burton – ICG – Oct12 – v1.1
1
Agenda
• Commodity Clusters
• Compute Pool
• Interconnects and Networks
• Shared Storage
• Login Layer and Middleware
HPC alternatives
HPC Facts• IBM Sequoia - Number 1 in top
500 with 1.572 million Cores.• 20 petaFLOPS ( Floating Point
Operations / Second ). Sciama 10 teraFLOPS.
• China has Number 5 plus 62 other systems in top 500, now ahead of Germany, UK, Japan and France.
• 75% of Top 500 use Intel processors.
Demystifying the Techno Babble
Demystifying the Techno Babble (2)
Commodity Clusters
7
Commodity Clusters• Made from commodity (off-the-shelf)
components (read PC’s).
• Consequently (relatively) cheap.
• Usually Linux based
• High availability storage (no single point of failure)
• Generic compute pool (cloned servers that can easily be replaced).
Cluster Concept
10
Compute Pool - Just a bunch of PC’s
In the “good-ol-days” things were simple ……….
11
In the “good-ol-days” things were simple ……….
12
… these days much more packed into the same space … but basically the
same!
13
These are the building blocks of HPC similar to Sciama
Total ICG Compute Pool > 1000 Cores
14
Coke or Pepsi – Chalk & Cheese
• Only 2 remaining commodity CPU makers are Intel and AMD.
• Latest AMD “Bulldozer” architecture competing with Intel “Sandy Bridge” Architecture.
• Both architectures are multi core ( Intel 48 cores)• Architectures use same memory / video cards /
hard drives etc• CPU speed constraints down to on-chip
transmission delays and heat dissipation ( 22nm ).
Intel-AMD – Bangs per Buck
Graphical Processing Units (GPU’s actually GPGPU’s – General purpose)
CPU’s still in charge
Special programminglanguage. CUDA and OpenCL
Three players:-IntelAMDNvidia
Cpu – multiple coresGpu – 100’s of cores
Interconnects and Networks
18
Interconnects and Networks
Moving away from the Processor towards the Internet you get slowerand slower due to Increased Latency and Reduced Bandwidth
Processor Interconnects
For processor running at 3.2GHz – QPI Bus would be 25GBytes / Second ( Kindle version of “War & Peace” is 2GBytes)
Peripheral Component Interconnect Express (PCIe)
PCI bus is interconnect to the outside world
External Networks
Interconnects are Parallel – Bytes / secondNetworks are serial – bits / second(shown here in B/s for comparison – eg. DAS is 6Gb/s)
Sciama Network
Sciama Traffic
Connecting to Sciama
26
Shared Storage
Raw Disks are Dumb
Remember: PATA, IDE (Advanced Technology Attachment)
Intelligence is in the File System
HPC’s Require many disks
Use High Capacity Arrays
HPC’s require large chucks of storage
Many RAID options 1-6 / 10 /50
Directly Attached Storage (DAS)
Directly Attached Storage (DAS)
Of limited use as cannot be shared.
Network Attached Storage (NAS)
NAS or Network Appliance
Network BW is often the bottleneck
NAS - Lustre File System
Lustre is an example of a distributed file system. There are many more.
Sometimes called a “Cluster” file system
NAS – Lustre
Often used with an Infiniband fabric.
Storage Area Networks
Storage Area Network using iSCSI
Fibre Channel High Availability SAN
Sciama NAS Storage
Sciama Storage Hardware
Storage is expensive.
250gbp / Tbytes
No Backup
Highly Available Hardware
Two paths to most components
Additional Sciama Storage/mnt/astro(1-5)
10 GbE
Login Layer and Middleware
48
Why Login Servers• Login servers will provide the gateway to the cluster.• Users can remotely login into the servers using “ssh” or
a Remote Desktop Client.• A desktop client gives a full working desktop in the
environment (can full screen)
Some users are at remote locations ..
50
Use of Remote Login Client
51
Executable and Jobscript setup in Login Layer
Jobs submitted to the queues
> qsub –q Queue1 run_script.sh
Scheduler prioritises and deems a job ready to run.
Job passed to resource manager
Resource manager (Torque) checks for available resources.
Job either runs in the compute pool or returns to queue