4
- Suren/Marco: Fully-functional backup server. Move katrin.kit.edu name to virtual machine
5
which will perform packet forwarding only. Or just VIP. Gluster vs. Master-Master MySQL replication
6
- Suren: How to proceed with student cluster computers? [ Enable Advanced/System Agent/iGPU. Find a
7
way to make snapshots (LFS vs BTRFs). User VMs using integrated video: KVM vs. OpenVZ/Docker/LXC ]
8
OpenSUSE 13.1 vs 13.3? Configuring the nodes, packages (remote boot? git? ...)
10
- Roman: Fix DFI (and ensure proper scallability)
11
- Matthias: Graph with bandwidth and latency of DirectGMA approach, DirectGMA vs. FPGA->CPU->GPU pathh
12
- Andrey: UFO-clustering benchmarks, scalability of ART methods on the cluster (Dresden cluster)
13
- Andrey: Why SBTV are performing badly on AMD platform? Scallability of SBTV
14
- Tomas: Laminography scallability on new AMD platform
15
- Timo: KIRO performance figures
17
- Suren: see details in bechmarking description
19
- Chuan: Performance on AMD, NVIDIA, Intel
20
* Why is not working on AMD/Intel? Use standard FFT instead of a full convolution. Or optimize it as in FBP?
21
- Chuan: Multi-GPU support, scallability
23
- Andrey: Implement Split-Bregman with framelets & framelets+tv in UFO.
24
- Andrey: Add DFI-based projectors (sinc)
25
- Andrey: Ensure scalling is working well on the cluster (Dresden)
26
- Andrey: More data sets (phase-contrast vs. absorbtion), simulation of Tomas (measure quality of algorithm)
27
- Andrey: Find if ART-reconstruction help for segmentation
29
- Nicholas: Stable version
30
- Nicholas: Zooming support
31
- Nicholas: Show high quality cuts when no user-interaction
32
- Nicholas: Full histogram as in version of Felix
33
- Nicholas: 4D visualization
34
- Andrey: Automatic background segmentation
35
- Student?: Advanced rendering: Data compression? DCT? Use texture tree's to optimize preloading.
37
- Andrey: Integration with slicer: show in low resolution / segment in high resolution.
42
- For the start in Jan we need proposals in July / August.
43
- For the start in May we need proposals in November.
49
+ We need reasons why Helmholtz will no pay this (science, not engineering)
50
+ Thinks best Engineering/Informatics
51
* Additional scientific stuff
52
+ Define metrics to compare quality, etc.
53
+ For phase-contrast, first perform FBP (ART), then, phase retrieval
54
* TV may be good, we only interested in the cell-countours in the embryo development
55
* Interested to see the example DFG-proposal we have applied
56
* Interested to participate in the PhD hiring process
58
- Andreas proposed Felix (GKSS)
63
- Nagios + Munin on ipepdvcompute* and cluster comoputers
64
- Add ipebessadei to Nagios. Failed KATRIN redirects are not detected
65
- Hardware: 2x 2.5'' to 3.5'' is taken by Danniel. Where is Raidsonic 2.5'' USB3 gehause + HDD (Chuan, Timo - order more)? HDD to WiFi was ordered? MiniPC complete (memory for Atom?)
66
Lost GPUBox IPs: ipepdvcompute1
67
192.168.26.161 just re-saved (IP is missing)
68
192.168.1.99 renamed to 192.168.26.162 (both IPs missing)
72
- darksoft.org expires on June 23, 2015
73
- On serv1, if glibc upgraded to 2.17 or 2.19, all applications linked with libm are
74
segfaulting. However, if we copy libm (only libm) from stage3-i486-20141209.tar.bz2,
75
everything seems to work. I have tried to configure and build myself disabling part
76
of optimizations, but to no avail (gcc 4.8.3 and latest kernel-headers).
77
Googling mention problem (or similar), but no solutions.
78
All stuff for future evaluation in /home/csa/glibc
83
- I'm waiting for info from Meik and David on the ankaimage-mnt configuration
85
- Check sterescopic view provided by ThreeJS
86
- Integrate automatic background removal
87
- Support segmented volumes
88
- Define multiple cache levels
89
- Define optimal storage format and storage engine
90
- Easy measurements of objects using pixel size on the web interface (voxel size is in meta-data)
91
- Point some specific function areas and export it as CSV file from web interface.
93
* Support ART in the cluster environment (transfer of object properties)
94
* RDMA based clustering
95
* On NVIDIA, common context for 9 GPUs significantly slow downs kernel execution penalty (from 20us to 150us).
96
We may look into using per-device contexts. Or spawn multiple UFO severs?
97
* UFO operations are either compiled on each invocation (very slow) or cached (only a single thread may use).
99
* Aysnchronous buffer API + overlapping.
100
* UFO/CUDA interoperability: OpenCL kernels can be compiled to PTX and further to nvidia cubins.
101
Support code should dublicated for CUDA and OpenCL.
103
* Currently there is no clFinish in the end if profilling is off. App will exit before GPUs are done
104
* In profilling charts show kernels and tasks using the same time scale (better interleave in single chart)
106
* Show transfer time in the kernels
108
- Optimizations and requirements of secondary filters
109
* ROI: Is currently always executed on CPU. For DFI this will increase the amount of
110
data transfered over PCIe significantly. We need hybrid version.
111
* FFT: Can we fix AMD library to fix performance on NVIDIA hardware and improve
114
* Optimize PyHST for Xeon Phi & GK110
115
* Check https://ispc.github.io/
116
* UFO: Migrate part of optimizations from PyHST (At least Kepler and GCN versions)
118
* UFO: Compute redundant coefficients using Hermetian symmetry
119
* Eliminate swapping quadrants stage before iFFT (just recompute coefficients in interpolation kernel)
120
* PyHST: Eleminate quandrant swapping and crop after iFFT (integrate it into the cudaMemcpy call splitting it
122
* In-place vs. out-of-place FFT? Currently, UFO is o-o-p, PyHST R2C is i-p and C2R is o-o-p.
124
- Compare quality and performance with Shao Lie application. Check our reconstruction with data
125
from Shao Lie and check Shao Lie reconstruction with data from Andrey's Article.
126
- Check if low-pass filtering may improve results on real-data.
127
- Provide projectors and back-projector in Matlab. Can we accelerate it like that? Multi-GPU.
128
- Number of kernel lauches takes significant time (2 launches per iteration per projection in SART) and
129
harms scallability on NVIDIA platform
130
- Fix running on clusters
131
- Split-bregmann (linearized-bregman), wavelets, overcomplete dictionaries
132
- TV-regularization on GPU
133
- Phase contrast and absorption contrast reconstruction and segmentation differs.
134
In phase contrast there are sharp edges. Find good dataset.
137
* in-place vs. out-of-place
139
- Background removal is most important now. Fighting ring artefacts, etc. Philipp algorithms.
140
- Positioning of object: where is head, where is tail, etc...
141
- There is inperfect ring-removal code fixing small artifacts (Tomas)
143
* Provide PyHST/OpenCL to deal with 8001 projections [Shall we?]
144
* Fix preprocessing SSE code of PyHST (alignment causes segmentations in some cases)
146
- Get update reconstruct script supporting reduced DFI output.
147
- Integrate DFI roi options to reconstruct script
153
- State recovery or storage of the current view (Felix)
154
- Settings storage (Transferfunction, Grayvalue selection) to apply on different volumes, maybe per user setting (Felix)
155
- Multiple time frames in zooming mode (Felix)
156
- Those orthogonal layers, which can be slid through the unvisible volume and on which the volume values are visible (Felix)
157
- Background removal (Philipp, Andrey)
158
- MapReduce with database
163
- UFO (with ASTOR) status and perspectives
164
- Reconstruction stack of UFO framework
165
* We might include phase contrast and grating interferometry (and possibly
166
other interesting preprocessing filters)
167
* We need to compare FBP and DFI quality. Noisy data? Undersampled data?
168
Is there cases when the FBP quality is superior?
169
* We need a better explanation why interpolation is limited to R (this is
170
not completely losseless) and that consequences it have for different types
171
of data (small details? Phantom used by Andrey?)
172
* UFO scalability issues. DFI is only add 50% performance with each GPU.
173
ART performance fully stops after 6th device.
174
* Thorsten Hopp have some expirience in defining metrics to compare reconstructed images
175
- Optimizing FBP for parallel architectures
177
* Verify FFT performance. Why iFFT 3 times slwoer when where is no significant difference
178
in FFT benchmark. Verify cuFFT C2R/R2C is no faster than C2C. Can we add it on the FFT chart
179
otherwise? Seems, cuFFT value for 2D FFT benchmark is wrong. Or is it due to C2C vs. C2R difference?
180
* In the stages benchmark, we have included H2D, but not D2H timing (in the null-task we doesn't
181
copy anything from device to host (we can force to copy, but it requires to set 'force-download'
182
property). Cropping is excluded from UFO-chain due to lack of GPU version. Cropping is
183
artificially excluded from PyHST numbers as well. It would be nice to include both
184
cropping stage and D2H transfer time in memory operations. The qundrant swapping should be
185
split in 2 (is it actually joined or more is skipped from post iFFT stages?): pre-iFFT
187
Actually what is included and excluded from performance chart?
195
- In OpenCL interpolate only half the volume by performing mirroring before 2D ifft
196
- Move zero-padding optimiztions to CUDA version
203
- Artur: Finish CRD modification and integrate in my version
204
- Suren: Return data from caches.
205
- Suren: Cache ADEI metadata in disconnected configurations
206
- Nicholas: Run-time interface with export of Raw Files (NetCDF for KITCube, free-format in general).
208
* Automatic data quality montiroing. Find trends and check if some sensors measurements
209
are fall out of the trends. Correlations, auto-correlations, etc. Also, introduce models,
212
We need to integrate more views from Weisser and provide them as secondary graphs
213
in ADEI (HATPRO, Lidar, Radar).
215
* ADEI: Filter 850 values among KATRIN temperatures
216
* ADEI: Support filter= parameter in get_data service
217
* ADEI: Provide access to the cached data instead of real (cache=3600) and use cache in
218
get_data.php by default
219
* ADEI: Handle ADEI archives with missing active database
220
* ADEI/Control: cache DB configuration and accept retries parameter (check for timeout error)
221
* ADEI: Check alternative D3 rendering engine
222
* ADEI: Temperature profile, etc. visualization in main graph
223
* ADEI: Download Manager for multi-purpose configurations
224
* ADEI: Download is failing for databases which source was cut, but cache stayed. No real data.
225
* ADEI: Database sorting as in configuration
226
* ADEI: JpGraph log axis are always starting from 1. Update jpgraph or configure manually
227
* ADEI: Notify in status that there is too many axes
229
* We may reconsider database selection. Mozilla is important. Just think about more clever way
230
to use indexedDb API.Also, something strange is happening with parsing data from cache (it stored
232
* Web sockets or HTTP Get?
233
* Workers in web threads
234
https://bugzilla.mozilla.org/show_bug.cgi?id=504553
236
- CRD: Prevent SELECT based aggregation for views (resolution should boldly define the results)
237
- Haag: Slow export (CSV) performance
238
- Haag: Joining of historical databases
239
- Andreas: Run/Subrun architecture (!)
240
- Andreas: Raw file exporter (!)
241
- Andreas: Mixed X & Y axes for Scatter plot, Problems with Log-scale labeling, Add cross-correlation plot
242
- ADEI autokiller to prevent hanged ADEI processes
249
- In pci cli while writting to files prevent telling (writted xx bytes) - quite mode
251
- Event Engine for HEB (clearing the padding), Payload Patch, Support for multiple devices
252
- Better camera locking
253
- We should have locking for the register access and camera start/stop/stream
254
operations. The DMA IP should remain for complete time of streaming.
255
However, after start we are unlocking camera and would like to allow
256
multiple threads to request frames. The stop should be called by the same
257
thread who executed start.
258
- Driver may report on process controlling DMA?
260
- Optimize buffer access, use hash-tables?
261
- Disconnect DMA and Event plugins...
262
- Support XML register models...
263
- Simplify driver, check VFIO...
264
- Allow DMA mapping of user-supplied memory...
265
- Support internal auto-triggering
266
- Properly use bus addresses
268
- Can we stabilize IB on ARM? Ask Pedraforca for help...
269
http://www.bsc.es/about-bsc/staff-directory/puzovic-nikola
270
http://www.bsc.es/computer-sciences/computer-architecture
271
http://webcache.googleusercontent.com/search?q=cache:NjZnHfkNqJ0J:on-demand.gputechconf.com/gtc/2013/presentations/S3064-Pedraforca-ARM-GPU-Cluster-HPC.pdf+&cd=1&hl=en&ct=clnk
272
- Can we do GPUDirect without Tesla?
273
- Play with remote GPUs (IB): manual and with MPI stack.
275
- Can we do direct Camera->IB transfers?
276
- GPUDirect for RDMA vs. GPUDirect for Video
277
- AMD DirectGMA? HSA specifications?
279
- Rewrite UCA driver efficiently
280
- Try to provide general rDMA based interface (?)
281
- GPI-2 replacement for MPI?
282
- Streaming to storage
283
- IO: Check Linus idea to use mmap I/O and page faults instead of O_DIRECT
288
- MultiGPU issues. Are we fine now?
292
- Control system (cFP interface with shared variable and DB, PVSS-based system, Direct cFP protocol)