3
- Camera: Improved performance, DirectGMA/GPUDirect, XML register model & python scripting.
4
RDMA is not yet integrated properly and even GPUDirect support for CMS is not even close to
5
be ready for production.
7
- UFO: Final report, paper in preparation. Final presentation about optimizations (may be
8
reported on better conference?) which can be probably integrated with GPUDirect paper.
10
- ADEI: New setup in Mexico. New conternerized version is almost ready.
12
- Infrastructure: New scalable and low effort infrastructure for data management setups in
19
- Drivers: Proper integration is time consuming and Michele is not collaborative. We also
20
have a proper DirectGMA version outside of the driver. Low latency GPUDirect has to be
21
implemented in the CUDA kernel and only slower version is possible.
22
* Generally, this I'd be interested to do, but probably it more wise to stop here until
23
there is a new project requiring high-bandwidth data streaming and online eellaboration.
24
* CMS version. GDRCopy is not stable and currenly limited to a single GPU page. We need
25
to re-implement for production. This relatively time consuming and probably will be solved
26
by NVIDIA if we wait. So, makes little sense unless we need it.
27
* Article: DirectGMA vs. GPUDirect, Event-based API vs Kernel-loop. Remote processing with
28
Infiniband (if this can be fixed). Generally, we have most of the stuff running, just need
29
to measure. May be as extension of CMS paper? Shall we add also tomography for small volumes?
30
We also can split over GPUs on z-axis. (though a good student would be helpful here.
31
* What are Matthias and Lorenzo doing, never reporting.
33
- Tomography: Help Andrey to write another paper summarizing that we have up to now. Should
34
add clustered execution benchmark. Results from Roman and Antonia. If possible, some results
35
illustrating the improvent of segmentation.
39
- Build PaaS infrastructure at IPE and recover student cluster. On top of it we can start
40
doing something for ADEI2. In general, this can be something on top of existing ADEI data
41
abstraction. Get data from ADEI caching tables and store it PostgreSQL/Cassandra and build
42
analysis toools. Jupiter notebooks could be interesting extension for data analysis on top
43
of existing ADEI. Further, it could be base for newer web display.
45
- Visualization. Push Nicolas to bring visualization to re-add all features of Felix
48
- Something new would be cool. Deep learning?
49
* Lighning search could be one application.
50
* Fully automatic segmentation of similar objects (there is efforts for automatic colorization
51
of black & white images and image classification which can be taken as an example):
52
http://tinyclouds.org/colorize/ http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf
53
http://machinelearningmastery.com/inspirational-applications-deep-learning/
54
This allows us to pre-generate the initialization for Philipp's algorithm.
55
* May be it can be applied for manuscripts of Claudio? We can easily produce large number
56
of samples to train the network.
57
* Automatic tagging of ADEI time series. Incremental feature learning from streamed data.
58
https://journalofbigdata.springeropen.com/articles/10.1186/s40537-014-0007-7
60
- Something completely new. Another ANKA to calloborate with?