|
1
|
- A project that should have changed the world, and a short introduction
into an exciting area of research
|
|
2
|
- A real life example would be trying to brush your teeth, read Mickey
Mouse and get dressed at the same time
- Hopefully by the time you get to school, you'll notice the hanger still
stuck in your jacket
- On a supercomputer: The Laplace PDE is solved on a polygon, in parallel
with an iterative Jacobi method
- Keep the communication between the nodes as short as possible! Otherwise,
the more processors needed, the slower...
|
|
3
|
- A set number (=p) of processors is defined before the program is run on
the supercomputer.
- OpenMP: The compiler takes over the whole parallelization task (the
programmer only has to give directives).
- MPI (Message Passing Interface): The programmer is the one who has to
define how the nodes communicate.
- But most of the time, a program is simply written, and is then started
multiple times with different initial parameters!!
|
|
4
|
- This laptop performs about a half million operations per second = ½
gigaflop; this is equal to the performance of a supercomputer 10 years
ago.
- A supercomputer with shared memory of about 470 gigaflops (HP Superdome)
- Linux Beowulf Cluster with distributed memory and 502 processors (every
process has its own memory) about 266 gigaflops
|
|
5
|
- The Japanese supercomputer Earth Simulator computes with 35 teraflops
(=35000 gigaflops)
- The Seti@home Project, the first successful grid-computing project
computes with 43 teraflops
|
|
6
|
- A CPU spends more than 80% of its time waiting for user input...
- Current operating systems could be running processes in the background,
without the user noticing anything
|
|
7
|
- If all the computers in the world were clustered, what kind of
performance could theoretically be achieved??
- 400 million computers @ ½ gigaflop = 200 million gigaflops = 200 000 teraflops
= approx. 2000 supercomputers!
|
|
8
|
|
|
9
|
- An old supercomputer distributes data from a radio-telescope to normal
computers.
- A small program installed on these computers analyzes the data in the
background.
- The analysis is done using a tedious Fourier transformation. The results
of the analysis are then sent back to the old supercomputer.
- Anyone can participate in the project! Even an 80486!
|
|
10
|
- Like Seti@home, United Devices distributes data sets, which are then
processed by normal computers in the background.
|
|
11
|
- In these two projects, Seti@home and Cancer Research, both running over
the Internet, the number of processors participating (=p) is not set in
advance... Computers can come and go as they wish.
- This is different from MPI and OpenMP!!
|
|
12
|
- Client / Server
- Examples: Seti@home, Cancer Research
- Peer-to-Peer network
- Examples: Kazaa, Gnutella, GPU, a @lobal processing unit??
|
|
13
|
|
|
14
|
- An idea that came from simulations done by a team at Princeton
University
|
|
15
|
- Computers connected to the Internet run GPU. GPU automatically connects
to a peer-to-peer network.
- GPU makes a scientific library available.
- Everyone, who has GPU installed, can also use other computers to perform
their own distributed calculations.
|
|
16
|
- GPU is has 3 parts.
- The routing layer forwards the calculation packets.
- A virtual machine interprets the calculation packets with the help of a
library of plugins.
- Plugins are compiled DLLs, which extend node functionality.
|
|
17
|
- Polish notation is introduced in order to simplify the virtual machine:
1 + 1 becomes 1,1,+.
- Calculation packets are disguised as file searches, e.g., a file search
for "GPU:1,1+" is interpreted as a calculation task.
|
|
18
|
- Two libraries (calculating pi and distributed calculation of the
discrete logarithm)
- Further development as an Open Source project failed due to the
complexity of the task
- You can download the prototype
from: http://sourceforge.net/projects/gpu
|
|
19
|
- Version 0.688 implemented everything described in the documentation...
but it is very unstable.
- Version 0.768 is very stable (thanks to the TGnutella components from
Kamil Pogorzelski). Results are not (yet) sent back.
|
|
20
|
|
|
21
|
|
|
22
|
- Global Grid Forum www.gridforum.org
- EU Grid: http://eu-datagrid.web.cern.ch
- Top 500 supercomputers: http://www.top500.org
- Seti@home http://setiathome.berkeley.edu
- Cancer Research http://members.ud.com/projects/cancer
|
|
23
|
|
Commentaires