Re: [Pytables-users] Chunk selection for optimized data access
Hi list, Il 05/06/2013 00:38, Anthony Scopatz ha scritto: On Tue, Jun 4, 2013 at 12:30 PM, Seref Arikan serefari...@gmail.com wrote: I think I've seen this in the release notes of 3.0. This is actually something that I'm looking into as well. So any experience/feedback about creating files in memory would be much appreciated. I think that you want to set parameters.DRIVER to H5DF_CORE [1]. I haven't ever used this personally, but it would be great to have an example script, if someone wants to write one ;) Be Well Anthony 1. http://pytables.github.io/usersguide/parameter_files.html#hdf5-driver-management thare is also a small example of usage in the cookbook [1] [1] http://pytables.github.io/cookbook/inmemory_hdf5_files.html ciao -- Antonio Valentino -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
You would be suprised to see how convenient HDF5 can be in small scale data :) There are cases where one may need to use binary serialization of a few thousand items, but still needing metadata, indexing and other nice features provided by HDF5/pyTables. On Wed, Jun 5, 2013 at 2:29 AM, Tim Burgess timburg...@mac.com wrote: I was playing around with in-memory HDF5 prior to the 3.0 release. Here's an example based on what I was doing. I looked over the docs and it does mention that there is an option to throw away the 'file' rather than write it to disk. Not sure how to do that and can't actually think of a use case where I would want to :-) And be wary, it is H5FD_CORE. On Jun 05, 2013, at 08:38 AM, Anthony Scopatz scop...@gmail.com wrote: I think that you want to set parameters.DRIVER to H5DF_CORE [1]. I haven't ever used this personally, but it would be great to have an example script, if someone wants to write one ;) import numpy as np import tables CHUNKY = 30 CHUNKX = 8640 if __name__ == '__main__': # create dataset and add global attrs file_path = 'demofile_chunk%sx%d.h5' % (CHUNKY, CHUNKX) with tables.open_file(file_path, 'w', title='PyTables HDF5 In-memory example', driver='H5FD_CORE') as h5f: # dummy some data lats = np.empty([4320]) lons = np.empty([8640]) # create some simple arrays lat_node = h5f.create_array('/', 'lat', lats, title='latitude') lon_node = h5f.create_array('/', 'lon', lons, title='longitude') # create a 365 x 4320 x 8640 CArray of 32bit float shape = (365, 4320, 8640) atom = tables.Float32Atom(dflt=np.nan) # chunk into daily slices and then further chunk days sst_node = h5f.create_carray(h5f.root, 'sst', atom, shape, chunkshape=(1, CHUNKY, CHUNKX)) # dummy up an ndarray sst = np.empty([4320, 8640], dtype=np.float32) sst.fill(30.0) # write ndarray to a 2D plane in the HDF5 sst_node[0] = sst -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
On 05.06.2013 10:31, Andreas Hilboll wrote: On 05.06.2013 03:29, Tim Burgess wrote: I was playing around with in-memory HDF5 prior to the 3.0 release. Here's an example based on what I was doing. I looked over the docs and it does mention that there is an option to throw away the 'file' rather than write it to disk. Not sure how to do that and can't actually think of a use case where I would want to :-) And be wary, it is H5FD_CORE. On Jun 05, 2013, at 08:38 AM, Anthony Scopatz scop...@gmail.com wrote: I think that you want to set parameters.DRIVER to H5DF_CORE [1]. I haven't ever used this personally, but it would be great to have an example script, if someone wants to write one ;) import numpy as np import tables CHUNKY = 30 CHUNKX = 8640 if __name__ == '__main__': # create dataset and add global attrs file_path = 'demofile_chunk%sx%d.h5' % (CHUNKY, CHUNKX) with tables.open_file(file_path, 'w', title='PyTables HDF5 In-memory example', driver='H5FD_CORE') as h5f: # dummy some data lats = np.empty([4320]) lons = np.empty([8640]) # create some simple arrays lat_node = h5f.create_array('/', 'lat', lats, title='latitude') lon_node = h5f.create_array('/', 'lon', lons, title='longitude') # create a 365 x 4320 x 8640 CArray of 32bit float shape = (365, 4320, 8640) atom = tables.Float32Atom(dflt=np.nan) # chunk into daily slices and then further chunk days sst_node = h5f.create_carray(h5f.root, 'sst', atom, shape, chunkshape=(1, CHUNKY, CHUNKX)) # dummy up an ndarray sst = np.empty([4320, 8640], dtype=np.float32) sst.fill(30.0) # write ndarray to a 2D plane in the HDF5 sst_node[0] = sst Thanks Tim, I adapted your example for my use case (I'm using the EArray class, because I need to continuously update my database), and it works well. However, when I use this with my own data (but also creating the arrays like you did), I'm running into errors like Could not wait on barrier. It seems like the HDF library is spawing several threads. Any idea what's going wrong? Can I somehow avoid HDF5 multithreading at runtime? Update: When setting max_blosc_threads=2 and max_numexpr_threads=2, everything seems to work as expected (but a bit on the slow side ...). With max_blosc_threads=4, the error pops up. Cheers, Andreas. -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
On 6/5/13 11:45 AM, Andreas Hilboll wrote: On 05.06.2013 10:31, Andreas Hilboll wrote: On 05.06.2013 03:29, Tim Burgess wrote: I was playing around with in-memory HDF5 prior to the 3.0 release. Here's an example based on what I was doing. I looked over the docs and it does mention that there is an option to throw away the 'file' rather than write it to disk. Not sure how to do that and can't actually think of a use case where I would want to :-) And be wary, it is H5FD_CORE. On Jun 05, 2013, at 08:38 AM, Anthony Scopatz scop...@gmail.com wrote: I think that you want to set parameters.DRIVER to H5DF_CORE [1]. I haven't ever used this personally, but it would be great to have an example script, if someone wants to write one ;) import numpy as np import tables CHUNKY = 30 CHUNKX = 8640 if __name__ == '__main__': # create dataset and add global attrs file_path = 'demofile_chunk%sx%d.h5' % (CHUNKY, CHUNKX) with tables.open_file(file_path, 'w', title='PyTables HDF5 In-memory example', driver='H5FD_CORE') as h5f: # dummy some data lats = np.empty([4320]) lons = np.empty([8640]) # create some simple arrays lat_node = h5f.create_array('/', 'lat', lats, title='latitude') lon_node = h5f.create_array('/', 'lon', lons, title='longitude') # create a 365 x 4320 x 8640 CArray of 32bit float shape = (365, 4320, 8640) atom = tables.Float32Atom(dflt=np.nan) # chunk into daily slices and then further chunk days sst_node = h5f.create_carray(h5f.root, 'sst', atom, shape, chunkshape=(1, CHUNKY, CHUNKX)) # dummy up an ndarray sst = np.empty([4320, 8640], dtype=np.float32) sst.fill(30.0) # write ndarray to a 2D plane in the HDF5 sst_node[0] = sst Thanks Tim, I adapted your example for my use case (I'm using the EArray class, because I need to continuously update my database), and it works well. However, when I use this with my own data (but also creating the arrays like you did), I'm running into errors like Could not wait on barrier. It seems like the HDF library is spawing several threads. Any idea what's going wrong? Can I somehow avoid HDF5 multithreading at runtime? Update: When setting max_blosc_threads=2 and max_numexpr_threads=2, everything seems to work as expected (but a bit on the slow side ...). With max_blosc_threads=4, the error pops up. Hmm, this seems like a bad interaction among threads in numexpr and blosc. I'm not sure why this is triggering because the libraries should execute at different times. Hmm is your app multi-threaded? Although Blosc has implemented a lock for preventing this situation in the latest releases, numexpr still lacks this protection. As the multithreading engine is the same for both packages, it should be relatively easy to implement the lock support to numexpr too. Volunteers? -- Francesc Alted -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
On 6/5/13 11:45 AM, Andreas Hilboll wrote: On 05.06.2013 10:31, Andreas Hilboll wrote: On 05.06.2013 03:29, Tim Burgess wrote: I was playing around with in-memory HDF5 prior to the 3.0 release. Here's an example based on what I was doing. I looked over the docs and it does mention that there is an option to throw away the 'file' rather than write it to disk. Not sure how to do that and can't actually think of a use case where I would want to :-) And be wary, it is H5FD_CORE. On Jun 05, 2013, at 08:38 AM, Anthony Scopatz scop...@gmail.com wrote: I think that you want to set parameters.DRIVER to H5DF_CORE [1]. I haven't ever used this personally, but it would be great to have an example script, if someone wants to write one ;) import numpy as np import tables CHUNKY = 30 CHUNKX = 8640 if __name__ == '__main__': # create dataset and add global attrs file_path = 'demofile_chunk%sx%d.h5' % (CHUNKY, CHUNKX) with tables.open_file(file_path, 'w', title='PyTables HDF5 In-memory example', driver='H5FD_CORE') as h5f: # dummy some data lats = np.empty([4320]) lons = np.empty([8640]) # create some simple arrays lat_node = h5f.create_array('/', 'lat', lats, title='latitude') lon_node = h5f.create_array('/', 'lon', lons, title='longitude') # create a 365 x 4320 x 8640 CArray of 32bit float shape = (365, 4320, 8640) atom = tables.Float32Atom(dflt=np.nan) # chunk into daily slices and then further chunk days sst_node = h5f.create_carray(h5f.root, 'sst', atom, shape, chunkshape=(1, CHUNKY, CHUNKX)) # dummy up an ndarray sst = np.empty([4320, 8640], dtype=np.float32) sst.fill(30.0) # write ndarray to a 2D plane in the HDF5 sst_node[0] = sst Thanks Tim, I adapted your example for my use case (I'm using the EArray class, because I need to continuously update my database), and it works well. However, when I use this with my own data (but also creating the arrays like you did), I'm running into errors like Could not wait on barrier. It seems like the HDF library is spawing several threads. Any idea what's going wrong? Can I somehow avoid HDF5 multithreading at runtime? Update: When setting max_blosc_threads=2 and max_numexpr_threads=2, everything seems to work as expected (but a bit on the slow side ...). BTW, can you really notice the difference between using 1, 2 or 4 threads? Can you show some figures? Just curious. -- Francesc Alted -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
Thanks Antonio and Tim! These are great. I think that one of these should definitely make it into the examples/ dir. Be Well Anthony On Wed, Jun 5, 2013 at 8:10 AM, Francesc Alted fal...@gmail.com wrote: On 6/5/13 11:45 AM, Andreas Hilboll wrote: On 05.06.2013 10:31, Andreas Hilboll wrote: On 05.06.2013 03:29, Tim Burgess wrote: I was playing around with in-memory HDF5 prior to the 3.0 release. Here's an example based on what I was doing. I looked over the docs and it does mention that there is an option to throw away the 'file' rather than write it to disk. Not sure how to do that and can't actually think of a use case where I would want to :-) And be wary, it is H5FD_CORE. On Jun 05, 2013, at 08:38 AM, Anthony Scopatz scop...@gmail.com wrote: I think that you want to set parameters.DRIVER to H5DF_CORE [1]. I haven't ever used this personally, but it would be great to have an example script, if someone wants to write one ;) import numpy as np import tables CHUNKY = 30 CHUNKX = 8640 if __name__ == '__main__': # create dataset and add global attrs file_path = 'demofile_chunk%sx%d.h5' % (CHUNKY, CHUNKX) with tables.open_file(file_path, 'w', title='PyTables HDF5 In-memory example', driver='H5FD_CORE') as h5f: # dummy some data lats = np.empty([4320]) lons = np.empty([8640]) # create some simple arrays lat_node = h5f.create_array('/', 'lat', lats, title='latitude') lon_node = h5f.create_array('/', 'lon', lons, title='longitude') # create a 365 x 4320 x 8640 CArray of 32bit float shape = (365, 4320, 8640) atom = tables.Float32Atom(dflt=np.nan) # chunk into daily slices and then further chunk days sst_node = h5f.create_carray(h5f.root, 'sst', atom, shape, chunkshape=(1, CHUNKY, CHUNKX)) # dummy up an ndarray sst = np.empty([4320, 8640], dtype=np.float32) sst.fill(30.0) # write ndarray to a 2D plane in the HDF5 sst_node[0] = sst Thanks Tim, I adapted your example for my use case (I'm using the EArray class, because I need to continuously update my database), and it works well. However, when I use this with my own data (but also creating the arrays like you did), I'm running into errors like Could not wait on barrier. It seems like the HDF library is spawing several threads. Any idea what's going wrong? Can I somehow avoid HDF5 multithreading at runtime? Update: When setting max_blosc_threads=2 and max_numexpr_threads=2, everything seems to work as expected (but a bit on the slow side ...). BTW, can you really notice the difference between using 1, 2 or 4 threads? Can you show some figures? Just curious. -- Francesc Alted -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
On Jun 06, 2013, at 04:19 AM, Anthony Scopatz scop...@gmail.com wrote:Thanks Antonio and Tim!These are great. I think that one of these should definitely make it into the examples/ dir.Be WellAnthonyOK. I have put up a pull request with the code added.https://github.com/PyTables/PyTables/pull/266Cheers, Tim -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
On 04.06.2013 05:35, Tim Burgess wrote: My thoughts are: - try it without any compression. Assuming 32 bit floats, your monthly 5760 x 2880 is only about 65MB. Uncompressed data may perform well and at the least it will give you a baseline to work from - and will help if you are investigating IO tuning. - I have found with CArray that the auto chunksize works fairly well. Experiment with that chunksize and with some chunksizes that you think are more appropriate (maybe temporal rather than spatial in your case). On Jun 03, 2013, at 10:45 PM, Andreas Hilboll li...@hilboll.de wrote: On 03.06.2013 14:43, Andreas Hilboll wrote: Hi, I'm storing large datasets (5760 x 2880 x ~150) in a compressed EArray (the last dimension represents time, and once per month there'll be one more 5760x2880 array to add to the end). Now, extracting timeseries at one index location is slow; e.g., for four indices, it takes several seconds: In [19]: idx = ((5000, 600, 800, 900), (1000, 2000, 500, 1)) In [20]: %time AA = np.vstack([_a[i,j] for i,j in zip(*idx)]) CPU times: user 4.31 s, sys: 0.07 s, total: 4.38 s Wall time: 7.17 s I have the feeling that this performance could be improved, but I'm not sure about how to properly use the `chunkshape` parameter in my case. Any help is greatly appreciated :) Cheers, Andreas. PS: If I could get significant performance gains by not using an EArray and therefore re-creating the whole database each month, then this would also be an option. -- Andreas. Thanks a lot, Anthony and Tim! I was able to get down the readout time considerably using chunkshape=(32, 32, 256) for my 5760x2880x150 array. Now, reading times are about as fast as I expected. the downside is that now, building up the database takes up a lot of time, because i get the data in chunks of 5760x2880x1. So I guess that writing the data to disk like this causes a load of IO operations ... My new question: Is there a way to create a file in-memory? If possible, I could then build up my database in-memory and then, once it's done, just copy the arrays to an on-disk file. Is that possible? If so, how? Thanks a lot for your help! -- Andreas. -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
I think I've seen this in the release notes of 3.0. This is actually something that I'm looking into as well. So any experience/feedback about creating files in memory would be much appreciated. Best regards Seref On Tue, Jun 4, 2013 at 2:09 PM, Andreas Hilboll li...@hilboll.de wrote: On 04.06.2013 05:35, Tim Burgess wrote: My thoughts are: - try it without any compression. Assuming 32 bit floats, your monthly 5760 x 2880 is only about 65MB. Uncompressed data may perform well and at the least it will give you a baseline to work from - and will help if you are investigating IO tuning. - I have found with CArray that the auto chunksize works fairly well. Experiment with that chunksize and with some chunksizes that you think are more appropriate (maybe temporal rather than spatial in your case). On Jun 03, 2013, at 10:45 PM, Andreas Hilboll li...@hilboll.de wrote: On 03.06.2013 14:43, Andreas Hilboll wrote: Hi, I'm storing large datasets (5760 x 2880 x ~150) in a compressed EArray (the last dimension represents time, and once per month there'll be one more 5760x2880 array to add to the end). Now, extracting timeseries at one index location is slow; e.g., for four indices, it takes several seconds: In [19]: idx = ((5000, 600, 800, 900), (1000, 2000, 500, 1)) In [20]: %time AA = np.vstack([_a[i,j] for i,j in zip(*idx)]) CPU times: user 4.31 s, sys: 0.07 s, total: 4.38 s Wall time: 7.17 s I have the feeling that this performance could be improved, but I'm not sure about how to properly use the `chunkshape` parameter in my case. Any help is greatly appreciated :) Cheers, Andreas. PS: If I could get significant performance gains by not using an EArray and therefore re-creating the whole database each month, then this would also be an option. -- Andreas. Thanks a lot, Anthony and Tim! I was able to get down the readout time considerably using chunkshape=(32, 32, 256) for my 5760x2880x150 array. Now, reading times are about as fast as I expected. the downside is that now, building up the database takes up a lot of time, because i get the data in chunks of 5760x2880x1. So I guess that writing the data to disk like this causes a load of IO operations ... My new question: Is there a way to create a file in-memory? If possible, I could then build up my database in-memory and then, once it's done, just copy the arrays to an on-disk file. Is that possible? If so, how? Thanks a lot for your help! -- Andreas. -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
I was playing around with in-memory HDF5 prior to the 3.0 release. Here's an example based on what I was doing.I looked over the docs and it does mention that there is an option to throw away the 'file' rather than write it to disk.Not sure how to do that and can't actually think of a use case where I would want to :-)And be wary, it is H5FD_CORE.On Jun 05, 2013, at 08:38 AM, Anthony Scopatz scop...@gmail.com wrote:I think that you want to set parameters.DRIVER to H5DF_CORE [1]. I haven't ever used this personally, but it would be great to have an example script, if someone wants to write one ;)import numpy as npimport tablesCHUNKY = 30CHUNKX = 8640if __name__ == '__main__': # create dataset and add global attrs file_path = 'demofile_chunk%sx%d.h5' % (CHUNKY, CHUNKX) with tables.open_file(file_path, 'w', title='PyTables HDF5 In-memory example', driver='H5FD_CORE') as h5f: # dummy some datalats = np.empty([4320])lons = np.empty([8640])# create some simple arrayslat_node = h5f.create_array('/', 'lat', lats, title='latitude')lon_node = h5f.create_array('/', 'lon', lons, title='longitude')# create a 365 x 4320 x 8640 CArray of 32bit floatshape = (365, 4320, 8640)atom = tables.Float32Atom(dflt=np.nan)# chunk into daily slices and then further chunk dayssst_node = h5f.create_carray(h5f.root, 'sst', atom, shape, chunkshape=(1, CHUNKY, CHUNKX))# dummy up an ndarraysst = np.empty([4320, 8640], dtype=np.float32)sst.fill(30.0)# write ndarray to a 2D plane in the HDF5sst_node[0] = sst-- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
[Pytables-users] Chunk selection for optimized data access
Hi, I'm storing large datasets (5760 x 2880 x ~150) in a compressed EArray (the last dimension represents time, and once per month there'll be one more 5760x2880 array to add to the end). Now, extracting timeseries at one index location is slow; e.g., for four indices, it takes several seconds: In [19]: idx = ((5000, 600, 800, 900), (1000, 2000, 500, 1)) In [20]: %time AA = np.vstack([_a[i,j] for i,j in zip(*idx)]) CPU times: user 4.31 s, sys: 0.07 s, total: 4.38 s Wall time: 7.17 s I have the feeling that this performance could be improved, but I'm not sure about how to properly use the `chunkshape` parameter in my case. Any help is greatly appreciated :) Cheers, Andreas. -- Get 100% visibility into Java/.NET code with AppDynamics Lite It's a free troubleshooting tool designed for production Get down to code-level detail for bottlenecks, with 2% overhead. Download for free and get started troubleshooting in minutes. http://p.sf.net/sfu/appdyn_d2d_ap2 ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
On 03.06.2013 14:43, Andreas Hilboll wrote: Hi, I'm storing large datasets (5760 x 2880 x ~150) in a compressed EArray (the last dimension represents time, and once per month there'll be one more 5760x2880 array to add to the end). Now, extracting timeseries at one index location is slow; e.g., for four indices, it takes several seconds: In [19]: idx = ((5000, 600, 800, 900), (1000, 2000, 500, 1)) In [20]: %time AA = np.vstack([_a[i,j] for i,j in zip(*idx)]) CPU times: user 4.31 s, sys: 0.07 s, total: 4.38 s Wall time: 7.17 s I have the feeling that this performance could be improved, but I'm not sure about how to properly use the `chunkshape` parameter in my case. Any help is greatly appreciated :) Cheers, Andreas. PS: If I could get significant performance gains by not using an EArray and therefore re-creating the whole database each month, then this would also be an option. -- Andreas. -- Get 100% visibility into Java/.NET code with AppDynamics Lite It's a free troubleshooting tool designed for production Get down to code-level detail for bottlenecks, with 2% overhead. Download for free and get started troubleshooting in minutes. http://p.sf.net/sfu/appdyn_d2d_ap2 ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
Hi Andreas, First off, nothing should be this bad, but What is the data type of the array? Also are you selecting chunksize manually or letting PyTables figure it out? Here are some things that you can try: 1. Query with fancy indexing, once. That is, rather than using a list comprehension just say, _a[zip(*idx)] 2. set _a.nrowsinbuf [1] to a much smaller value (1, 5, or 10) which is more appropriate for pulling out individual indexes. Lastly, it is my opinion that the iteration mechanics are slower than they can / should be. I have a bunch of ideas about how to make them faster AND clean up the code base but I won't have a ton of time to work on them in the near term. However, if this is something that you are interested in, that would be great! I'd love to help out anyone who was willing to take this on. Be Well Anthony 1. http://pytables.github.io/usersguide/libref/hierarchy_classes.html#tables.Leaf.nrowsinbuf On Mon, Jun 3, 2013 at 7:45 AM, Andreas Hilboll li...@hilboll.de wrote: On 03.06.2013 14:43, Andreas Hilboll wrote: Hi, I'm storing large datasets (5760 x 2880 x ~150) in a compressed EArray (the last dimension represents time, and once per month there'll be one more 5760x2880 array to add to the end). Now, extracting timeseries at one index location is slow; e.g., for four indices, it takes several seconds: In [19]: idx = ((5000, 600, 800, 900), (1000, 2000, 500, 1)) In [20]: %time AA = np.vstack([_a[i,j] for i,j in zip(*idx)]) CPU times: user 4.31 s, sys: 0.07 s, total: 4.38 s Wall time: 7.17 s I have the feeling that this performance could be improved, but I'm not sure about how to properly use the `chunkshape` parameter in my case. Any help is greatly appreciated :) Cheers, Andreas. PS: If I could get significant performance gains by not using an EArray and therefore re-creating the whole database each month, then this would also be an option. -- Andreas. -- Get 100% visibility into Java/.NET code with AppDynamics Lite It's a free troubleshooting tool designed for production Get down to code-level detail for bottlenecks, with 2% overhead. Download for free and get started troubleshooting in minutes. http://p.sf.net/sfu/appdyn_d2d_ap2 ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users -- Get 100% visibility into Java/.NET code with AppDynamics Lite It's a free troubleshooting tool designed for production Get down to code-level detail for bottlenecks, with 2% overhead. Download for free and get started troubleshooting in minutes. http://p.sf.net/sfu/appdyn_d2d_ap2___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
My thoughts are:- try it without any compression. Assuming 32 bit floats, your monthly 5760 x 2880 is only about 65MB. Uncompressed data may perform well and at the least it will give you a baseline to work from - and will help if you are investigating IO tuning.- I have found with CArray that the auto chunksize works fairly well. Experiment with that chunksize and with some chunksizes that you think are more appropriate (maybe temporal rather than spatial in your case).On Jun 03, 2013, at 10:45 PM, Andreas Hilboll li...@hilboll.de wrote:On 03.06.2013 14:43, Andreas Hilboll wrote: Hi,I'm storing large datasets (5760 x 2880 x ~150) in a compressed EArray (the last dimension represents time, and once per month there'll be one more 5760x2880 array to add to the end).Now, extracting timeseries at one index location is slow; e.g., for four indices, it takes several seconds:In [19]: idx = ((5000, 600, 800, 900), (1000, 2000, 500, 1))In [20]: %time AA = np.vstack([_a[i,j] for i,j in zip(*idx)]) CPU times: user 4.31 s, sys: 0.07 s, total: 4.38 s Wall time: 7.17 sI have the feeling that this performance could be improved, but I'm not sure about how to properly use the `chunkshape` parameter in my case.Any help is greatly appreciated :)Cheers, Andreas. PS: If I could get significant performance gains by not using an EArray and therefore re-creating the whole database each month, then this would also be an option. -- Andreas. -- Get 100% visibility into Java/.NET code with AppDynamics Lite It's a free troubleshooting tool designed for production Get down to code-level detail for bottlenecks, with 2% overhead. Download for free and get started troubleshooting in minutes. http://p.sf.net/sfu/appdyn_d2d_ap2 ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users-- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
Opps! I forgot to mention CArray! On Mon, Jun 3, 2013 at 10:35 PM, Tim Burgess timburg...@mac.com wrote: My thoughts are: - try it without any compression. Assuming 32 bit floats, your monthly 5760 x 2880 is only about 65MB. Uncompressed data may perform well and at the least it will give you a baseline to work from - and will help if you are investigating IO tuning. - I have found with CArray that the auto chunksize works fairly well. Experiment with that chunksize and with some chunksizes that you think are more appropriate (maybe temporal rather than spatial in your case). On Jun 03, 2013, at 10:45 PM, Andreas Hilboll li...@hilboll.de wrote: On 03.06.2013 14:43, Andreas Hilboll wrote: Hi, I'm storing large datasets (5760 x 2880 x ~150) in a compressed EArray (the last dimension represents time, and once per month there'll be one more 5760x2880 array to add to the end). Now, extracting timeseries at one index location is slow; e.g., for four indices, it takes several seconds: In [19]: idx = ((5000, 600, 800, 900), (1000, 2000, 500, 1)) In [20]: %time AA = np.vstack([_a[i,j] for i,j in zip(*idx)]) CPU times: user 4.31 s, sys: 0.07 s, total: 4.38 s Wall time: 7.17 s I have the feeling that this performance could be improved, but I'm not sure about how to properly use the `chunkshape` parameter in my case. Any help is greatly appreciated :) Cheers, Andreas. PS: If I could get significant performance gains by not using an EArray and therefore re-creating the whole database each month, then this would also be an option. -- Andreas. -- Get 100% visibility into Java/.NET code with AppDynamics Lite It's a free troubleshooting tool designed for production Get down to code-level detail for bottlenecks, with 2% overhead. Download for free and get started troubleshooting in minutes. http://p.sf.net/sfu/appdyn_d2d_ap2 ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users -- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
Re: [Pytables-users] Chunk selection for optimized data access
and for the record...yes, it should be much faster than 4 seconds. foo = np.empty([5760,2880,150],dtype=np.float32) idx = ((5000,600,800,900),(1000,2000,500,1)) import time t0 = time.time();bar=np.vstack([foo[i,j] for i,j in zip(*idx)]);t1=time.time(); print t1-t00.000144004821777On Jun 03, 2013, at 10:45 PM, Andreas Hilboll li...@hilboll.de wrote:On 03.06.2013 14:43, Andreas Hilboll wrote: Hi,I'm storing large datasets (5760 x 2880 x ~150) in a compressed EArray (the last dimension represents time, and once per month there'll be one more 5760x2880 array to add to the end).Now, extracting timeseries at one index location is slow; e.g., for four indices, it takes several seconds:In [19]: idx = ((5000, 600, 800, 900), (1000, 2000, 500, 1))In [20]: %time AA = np.vstack([_a[i,j] for i,j in zip(*idx)]) CPU times: user 4.31 s, sys: 0.07 s, total: 4.38 s Wall time: 7.17 sI have the feeling that this performance could be improved, but I'm not sure about how to properly use the `chunkshape` parameter in my case.Any help is greatly appreciated :)Cheers, Andreas. PS: If I could get significant performance gains by not using an EArray and therefore re-creating the whole database each month, then this would also be an option. -- Andreas. -- Get 100% visibility into Java/.NET code with AppDynamics Lite It's a free troubleshooting tool designed for production Get down to code-level detail for bottlenecks, with 2% overhead. Download for free and get started troubleshooting in minutes. http://p.sf.net/sfu/appdyn_d2d_ap2 ___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users-- How ServiceNow helps IT people transform IT departments: 1. A cloud service to automate IT design, transition and operations 2. Dashboards that offer high-level views of enterprise services 3. A single system of record for all IT processes http://p.sf.net/sfu/servicenow-d2d-j___ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users