Back to the main page.

Bug 1667 - improve the caching implementation in ft_read_data

Status ASSIGNED
Reported 2012-08-22 09:50:00 +0200
Modified 2015-01-08 10:09:24 +0100
Product: FieldTrip
Component: realtime
Version: unspecified
Hardware: PC
Operating System: Mac OS
Importance: P3 normal
Assigned to: Jörn M. Horschig
URL:
Tags:
Depends on:
Blocks:
See also: http://bugzilla.fcdonders.nl/show_bug.cgi?id=1453

Robert Oostenveld - 2012-08-22 09:50:04 +0200

reading the manuscript from Wouter, it occurred to me that he probably did not use the local caching that is implemented in ft_read_data, at least he does not mention it. It might speed things up. This might also be of interest to brainstream. Around line 191, together with the section at line 1088, there now is something like if cache=yes try to get data from cache if succes return else get data from the file or FT buffer store in cache for the next time end end It needs to be improved w.r.t. the total cache size (it now will grow out of hand), i.e. old data should be flushed. Furthermore, the cache success is now boolean (yes/no), a partial cache hit is not supported. That means that the overlapping segment reading as used by Wouter would not benefit from this. If ft_fetch_data were made smarter (or perhaps it can already do it) and would return nans for the missing part of the data upon a partial cache hit, the subsequent read of the buffer/file could be restricted to only the part that was not present in the cache. I.e. something like dat = ft_fetch_data(...) if any(isnan(dat(1,;)) % get the section or sections of data that were not in the cache begsample = ... endsample = ... read the data from the actual stream/file, use a recursive ft_read_data call with caching switched off add the new segments to the cache fill in the missing pieces of the requested data return end I think it would be easiest to implement if the two sections now (at 191 and 1088) would be merged into one, using a recursive call to ft_read_data


Robert Oostenveld - 2012-08-22 09:59:24 +0200

another request that is clear from Wouters report is that it is desired to have WAIT_GET_DAT and integrate the blocking into the ft_read_data call. See also bug #1428 and http://fieldtrip.fcdonders.nl/development/realtime/buffer_protocol#suggested_extensions_to_the_protocol


Jörn M. Horschig - 2012-08-22 10:14:33 +0200

Good idea! Indeed, he did not use caching. It will make using the buffer more flexible and ease using it for real-time feedback application. I'm gonna work on that after the BioMag conference


Robert Oostenveld - 2012-08-22 10:25:44 +0200

(In reply to comment #2) please coordinate with Boris to ensure that the right stuff ends up in either the low-level C reference implementation or in the high-level MATLAB implementation.


Jörn M. Horschig - 2012-09-20 13:23:42 +0200

As a sidenote: you can ask the buffer for the amount of samples which are available and thus retrieve the last x samples. This invalidates the try-catch approach that Wouter used, which initiated this improvement. So one thing, which is theeasy fix, is to improve the caching behaviour, so that you can ask for retrieving data from the cache, and if not available, only the data that is missing will be pulled over from the buffer. The other thing concerns a new command for the buffer, so that the buffer can wait for the requested data to be available rather than returning an error. I guess this second one (WAIT_GET_DAT) would need to go into the buffer.c file, right, while the first one would be alright in the higher-level matlab function. Boris, any thoughts on this or additions?


Boris Reuderink - 2012-09-25 10:52:19 +0200

One concern that I have is that Wouter used an approach that is probably sub-optimal for what was already available. For me this indicates that for each solution, we should think of how we inform and/or educate our users. @Jörn: shall we discuss this tomorrow at 10:00?


Jörn M. Horschig - 2012-09-25 11:13:18 +0200

yes, let's discuss it, 10 am is fine. Got a group meeting at 11 though.


Jörn M. Horschig - 2012-09-26 15:30:25 +0200

tasks: ft_read_data - change cachedata.trl to .sampleinfo to be consistent with current guidelines - add a userspecific cfg option for cachelength, specified in samples - remove trials if samples > thresholds (only remove full trials, not partially) profiling/buffer: - set up a buffer and use qsubcellfun to simulate many processes requesting events/header from the buffer. check timing - test how fast pulling data over from the buffer is in the MEG w/o caching profiling/ft_fetch_data - write a test script to check speed of ft_fetch_data when inputdata has chunks of a few 100ms and data is requested that extends over several chunks (including asking for data outside the chunks). If time > (see above), think about if function can be faster or whether caching in ft_read_data does make sense.


Jörn M. Horschig - 2012-09-27 10:33:35 +0200

insights: 1) caching could not be used before, because ft_fetch_data works in .sampleinfo rather than on .trl. changed this accordingly in ft_read_data 2) ft_fetch_data cannot deal with requesting data from outside the input data. fixed this. 3) 100 times fetching a 4s chunk of data from 20 cached segments takes about 1.5seconds, from 100 segments it takes about 2 seconds, with 1000segments about 5.5econds. Remember that this is for 100 iterations, so even with 1000segments it takes only about 50ms per fetch, with 20 cached segements one fetch takes on average 15ms. The factor that is most determined is the length of the requested data, independently of the length of chunks in the cache, e.g when fetching only 1s time decreases by 25-50% (depending on cachesize). Remains the question how that compares to getting all data from the buffer compared to only a small part. 4) ft_fetch_data could not deal properly with only 1 trial, i.e. it did not return nans but errored when data outside the trial was requested. Fixed this so that it is consistent with many trials. Made a testscript for this special case. http://code.google.com/p/fieldtrip/source/detail?r=6554


Robert Oostenveld - 2012-09-27 11:12:46 +0200

(In reply to comment #8) regarding item 3: This suggests that the ft_fetch_data function is memory inefficient, i.e. moving more bytes around in memory than it should be doing. Have you done profile clear; profile on ... run script profile report to see the lines of code that take the most time?


Jörn M. Horschig - 2012-09-27 12:07:20 +0200

about 70% is due to ft_checkdata. once it is done by ft_fetch_header which an easily be saved by fetching the header in advance. the other call to ft_checkdata is in ft_fetch_data. In ft_checkdata, most time (90%) consumed by ft_datatype_raw and here in particular fixtimeaxis (99%), line 211 takes up 60% of this: endtime(i) = data.time{i}(end); the rest is consumed mostly (50%) by looping through trials (~15-20%): for xlop=1:length(utrl) ok = trialnum==utrl(xlop); smps = samplenum(ok); dat(:,ok) = data.trial{utrl(xlop)}(chanindx,smps); end for trllop=1:trlnum trllen(trllop) = size(data.trial{trllop},2); end or by dealing with matrices when having many trials (i.e. many samples): count = zeros(1, maxsample, 'int32'); trialnum = zeros(1, maxsample, 'int32'); samplenum = zeros(1, maxsample, 'int32'); dat = NaN(numel(chanindx),endsample-begsample+1);


Roemer van der Meij - 2012-09-28 14:48:13 +0200

Hey Jorn, your added code crashed if there was an offset present in the trl, I had a quick look, should be fixed as of now.


Philip van den Broek - 2012-10-08 13:25:05 +0200

Hi Jörn, As a suggestion, in the current implementation it seems like caching only seems to work if only one buffer is involved. Perhaps, you're already implementing such feature, but if not it might be worthwhile to consider making the caching "buffer-address" dependent. Gr. Philip


Robert Oostenveld - 2012-10-08 15:54:51 +0200

(In reply to comment #12) or more general, if dat1 = ft_read_data(datafile1) dat2 = ft_read_data(datafile2) are called with different datafiles, then ft_read_data should cache both and not get confused.


Jörn M. Horschig - 2014-01-29 14:37:21 +0100

caching from multiple streams might be tricky, because one would need a unique way to identify each stream. Right now, caching is implemented by comparing the channel labels of the cached and the input data. I would suggest to make a persistent variable called "dataset" which is a cell-array and contains as many entries as cached datasets (similar for cache_data). dataset will store the unique name of the dataset and only if that matches the already stored data, it will try to fetch data from the cached data structure another problem, and potential bug, is that if cached data is fetched, everything outside the current cached data is filled with nans. the question is, in what cases should it be kept like this and when should the nans be tried to filled current data?


Jörn M. Horschig - 2015-01-07 11:04:30 +0100

(In reply to Jörn M. Horschig from comment #14) the nan-fill question still needs to be addressed by one of the senseis ;)


Robert Oostenveld - 2015-01-08 10:02:26 +0100

(In reply to Jörn M. Horschig from comment #15) In order to cache data from multiple datasets it needs to keep track of the data source (i.e. string like "localhost://buffer:1972" and of the channels that were read previously from that source. Would this work by means of having two cell array of structs like cachefilename{i} = string, i.e. filename input arg and cachedata{i}.label = hdr.label(chanindx); cachedata{i}.fsample = hdr.Fs; cachedata{i}.time = {}; cachedata{i}.trial = {}; cachedata{i}.cfg = []; % does not seem to be used, so pleas remove cachedata{i}.sampleinfo = zeros(0,2); where the 2nd must be compatible with ft_datatype_raw? To me this seems like a perfectly fine implementation. Consider the following unusual test cases with caching for i=1:10 % read from 10 sources which have one channel each dat(i,:) = ft_read_data(dataset{i}, 'chanindx', 1); end for i=1:10 % read from 1 source which has 10 channels dat(i,:) = ft_read_data(dataset, 'chanindx', i); end For me it is now clear how cache filename and cache data would look like. Please go ahead and implement like this.


Jörn M. Horschig - 2015-01-08 10:06:14 +0100

alright, sounds like a good plan. what about the nan problem? currently, if you ask for data that is already partly cached, the currently not chached part will be filled with nans. It could be filled by reading from the original dataset (which could sometimes fail). Should this be a user-choice?


Robert Oostenveld - 2015-01-08 10:09:24 +0100

(In reply to Jörn M. Horschig from comment #17) (In reply to Jörn M. Horschig from comment #15) regarding the nan filling. I suggest to change try dat = ft_fetch_data(cachedata, 'begsample', begsample', 'endsample', endsample); % fprintf('caching succeeded\n'); return catch % fprintf('caching failed\n'); end into try dat = ft_fetch_data(cachedata, 'begsample', begsample', 'endsample', endsample); assert(~any(isnan(dat(:))); % fprintf('caching succeeded\n'); return catch % fprintf('caching failed\n'); end where the assert will error if there is a nan and subsequently a new read operation will be performed. Thie results in the most recent data being read. If that data still contains a nan (just like the cached data), then there is nothing we can do about it.