Tall Arrays for Out-of-Memory Data - MATLAB & Simulink - MathWorks Switzerland (2024)

What is a Tall Array?

Since the data is not loaded into memory all at once, tall arrays can be arbitrarily large in the first dimension (that is, they can have any number of rows). Instead of writing special code that takes into account the huge size of the data, such as with techniques like MapReduce, tall arrays let you work with large data sets in an intuitive manner that is similar to the way you would work with in-memory MATLAB^® arrays. Many core operators and functions work the same with tall arrays as they do with in-memory arrays. MATLAB works with small blocks of the data at a time, handling all of the data chunking and processing in the background, so that common expressions, such as A+B, work with big data sets.

Benefits of Tall Arrays

Unlike in-memory arrays, tall arrays typically remain unevaluated until you request that the calculations be performed using the gather function. This lazy evaluation allows you to work quickly with large data sets. When you eventually request output using gather, MATLAB combines the queued calculations where possible and takes the minimum number of passes through the data. The number of passes through the data greatly affects execution time, so it is recommended that you request output only when necessary.

Creating Tall Tables

Tall tables are like in-memory MATLAB tables, except that they can have any number of rows. To create a tall table from a large data set, you first need to create a datastore for the data. If the datastore ds contains tabular data, then tall(ds) returns a tall table or tall timetable containing the data. See Datastore for more information about creating datastores.

Create a spreadsheet datastore that points to a tabular file of airline flight data. For folders that contain a collection of files, you can specify the entire folder location, or use the wildcard character, '*.csv', to include multiple files with the same file extension in the datastore. Clean the data by treating 'NA' values as missing data so that tabularTextDatastore replaces them with NaN values. Also, set the format of a few text variables to %s so that tabularTextDatastore reads them as cell arrays of character vectors.

ds = tabularTextDatastore('airlinesmall.csv');ds.TreatAsMissing = 'NA';ds.SelectedFormats{strcmp(ds.SelectedVariableNames,'TailNum')} = '%s';ds.SelectedFormats{strcmp(ds.SelectedVariableNames,'CancellationCode')} = '%s';

Create a tall table from the datastore. When you perform calculations on this tall table, the underlying datastore reads blocks of data and passes them to the tall table to process. Neither the datastore nor the tall table retain any of the underlying data.

tt = tall(ds)

tt = M×29 tall table Year Month DayofMonth DayOfWeek DepTime CRSDepTime ArrTime CRSArrTime UniqueCarrier FlightNum TailNum ActualElapsedTime CRSElapsedTime AirTime ArrDelay DepDelay Origin Dest Distance TaxiIn TaxiOut Cancelled CancellationCode Diverted CarrierDelay WeatherDelay NASDelay SecurityDelay LateAircraftDelay ____ _____ __________ _________ _______ __________ _______ __________ _____________ _________ _______ _________________ ______________ _______ ________ ________ ______ _____ ________ ______ _______ _________ ________________ ________ ____________ ____________ ________ _____________ _________________ 1987 10 21 3 642 630 735 727 'PS' 1503 'NA' 53 57 NaN 8 12 'LAX' 'SJC' 308 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 1987 10 26 1 1021 1020 1124 1116 'PS' 1550 'NA' 63 56 NaN 8 1 'SJC' 'BUR' 296 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 1987 10 23 5 2055 2035 2218 2157 'PS' 1589 'NA' 83 82 NaN 21 20 'SAN' 'SMF' 480 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 1987 10 23 5 1332 1320 1431 1418 'PS' 1655 'NA' 59 58 NaN 13 12 'BUR' 'SJC' 296 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 1987 10 22 4 629 630 746 742 'PS' 1702 'NA' 77 72 NaN 4 -1 'SMF' 'LAX' 373 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 1987 10 28 3 1446 1343 1547 1448 'PS' 1729 'NA' 61 65 NaN 59 63 'LAX' 'SJC' 308 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 1987 10 8 4 928 930 1052 1049 'PS' 1763 'NA' 84 79 NaN 3 -2 'SAN' 'SFO' 447 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 1987 10 10 6 859 900 1134 1123 'PS' 1800 'NA' 155 143 NaN 11 -1 'SEA' 'LAX' 954 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :

The display indicates that the number of rows, M,is currently unknown. MATLAB displays some of the rows, and thevertical ellipses : indicate that more rows existin the tall table that are not currently being displayed.

Creating Tall Timetables

If the data you are working with has a time associated with each row of data, then you can use a tall timetable to work on the data. For information on creating tall timetables, see Extended Capabilities (timetable).

In this case, the tall table tt has times associated with each row, but they are broken down into several table variables such as Year, Month, DayofMonth, and so on. Combine all of these pieces of datetime information into a single new tall datetime variable Dates, which is based on the departure times DepTime. Then, create a tall timetable using Dates as the row times. Since Dates is the only datetime variable in the table, the table2timetable function automatically uses it for the row times.

hrs = (tt.DepTime - mod(tt.DepTime,100))/100;mins = mod(tt.DepTime,100);tt.Dates = datetime(tt.Year, tt.Month, tt.DayofMonth, hrs, mins, 0);tt(:,1:8) = [];TT = table2timetable(tt)

TT = M×21 tall timetable Dates UniqueCarrier FlightNum TailNum ActualElapsedTime CRSElapsedTime AirTime ArrDelay DepDelay Origin Dest Distance TaxiIn TaxiOut Cancelled CancellationCode Diverted CarrierDelay WeatherDelay NASDelay SecurityDelay LateAircraftDelay ____________________ _____________ _________ _______ _________________ ______________ _______ ________ ________ ______ _____ ________ ______ _______ _________ ________________ ________ ____________ ____________ ________ _____________ _________________ 21-Oct-1987 06:42:00 'PS' 1503 'NA' 53 57 NaN 8 12 'LAX' 'SJC' 308 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 26-Oct-1987 10:21:00 'PS' 1550 'NA' 63 56 NaN 8 1 'SJC' 'BUR' 296 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 23-Oct-1987 20:55:00 'PS' 1589 'NA' 83 82 NaN 21 20 'SAN' 'SMF' 480 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 23-Oct-1987 13:32:00 'PS' 1655 'NA' 59 58 NaN 13 12 'BUR' 'SJC' 296 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 22-Oct-1987 06:29:00 'PS' 1702 'NA' 77 72 NaN 4 -1 'SMF' 'LAX' 373 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 28-Oct-1987 14:46:00 'PS' 1729 'NA' 61 65 NaN 59 63 'LAX' 'SJC' 308 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 08-Oct-1987 09:28:00 'PS' 1763 'NA' 84 79 NaN 3 -2 'SAN' 'SFO' 447 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN 10-Oct-1987 08:59:00 'PS' 1800 'NA' 155 143 NaN 11 -1 'SEA' 'LAX' 954 NaN NaN 0 'NA' 0 NaN NaN NaN NaN NaN : : : : : : : : : : : : : : : : : : : : : : : : 

Creating Tall Arrays

When you extract a variable from a tall table or tall timetable, the result is a tall array of the appropriate underlying data type. A tall array can be a numeric, logical, datetime, duration, calendar duration, categorical, string, or cell array. Also, you can convert an in-memory array A into a tall array with tA = tall(A). The in-memory array A must have one of the supported data types.

Extract the arrival delay ArrDelay from thetall timetable TT. This creates a new tall arrayvariable with underlying data type double.

a = TT.ArrDelay

a = M×1 tall double column vector 8 8 21 13 4 59 3 11 : :

The classUnderlying and isaUnderlying functions are useful todetermine the underlying data type of a tall array.

Lazy Evaluation

One important aspect of tall arrays is that as you work with them, most operations are not performed immediately. These operations appear to execute quickly, because the actual computation is deferred until you specifically request that the calculations be performed. You can trigger evaluation of a tall array with either the gather function (to bring the result into memory) or the write function (to write the result to disk). This lazy evaluation is important because even a simple command like size(X) executed on a tall array with a billion rows is not a quick calculation.

As you work with tall arrays, MATLAB keeps track of allof the operations to be carried out. This information is then usedto optimize the number of passes through the data that will be requiredwhen you request output with the gather function.Thus, it is normal to work with unevaluated tall arrays and requestoutput only when you require it. For more information, see Lazy Evaluation of Tall Arrays.

Calculate the mean and standard deviation of the arrival delay.Use these values to construct the upper and lower thresholds for delaysthat are within one standard deviation of the mean. Notice that theresult of each operation indicates that the array has not been calculatedyet.

m = mean(a,'omitnan')

m = tall double ?Preview deferred. Learn more.

s = std(a,'omitnan')

s = tall ?Preview deferred. Learn more.

one_sigma_bounds = [m-s m m+s]

one_sigma_bounds = M×N×... tall array ? ? ? ... ? ? ? ... ? ? ? ... : : : : : :Preview deferred. Learn more.

Evaluation with gather

The benefit of lazy evaluation is that when the time comes for MATLAB to perform the calculations, it is often possible to combine the operations in such a way that the number of passes through the data is minimized. So even if you perform many operations, MATLAB only makes extra passes through the data when absolutely necessary.

The gather function forces evaluation of all queued operations and brings the resulting output into memory. For this reason, you can think of gather as a bridge between tall arrays and in-memory arrays. For example, you cannot control if or while loops using a tall logical array, but once the array is evaluated with gather it becomes an in-memory logical array that you can use in these contexts.

Since gather returns the entire result in MATLAB, you should make sure that the result will fit in memory.

Use gather to calculate one_sigma_bounds andbring the result into memory. In this case, one_sigma_bounds requiresseveral operations to calculate, but MATLAB combines the operationsinto one pass through the data. Since the data in this example issmall, gather executes quickly. However, theelimination of passes through the data becomes more valuable as thesize of your data increases.

sig1 = gather(one_sigma_bounds)

Evaluating tall expression using the Local MATLAB Session:- Pass 1 of 1: Completed in 1.5 secEvaluation completed in 1.8 secsig1 = -23.4572 7.1201 37.6975

You can specify multiple inputs and outputs to gather ifyou want to evaluate several tall arrays at once. This technique isfaster than calling gather multiple times. Forexample, calculate the minimum and maximum arrival delay. Computedseparately, each value requires a pass through the data to calculatefor a total of two passes. However, computing both values simultaneouslyrequires only one pass through the data.

[max_delay, min_delay] = gather(max(a),min(a))

Evaluating tall expression using the Local MATLAB Session:- Pass 1 of 1: Completed in 1.1 secEvaluation completed in 1.1 secmax_delay = 1014min_delay = -64

These results indicate that on average, most flights arriveabout 7 minutes late. But it is within one standard deviation fora flight to be up to 37 minutes late or 23 minutes early. The quickestflight in the data set arrived about an hour early, and the latestflight was delayed by many hours.

Saving, Loading, and Checkpointing Tall Arrays

The save function savesthe state of a tall array, but does not copyany of the data. The resulting .mat file is typicallysmall. However, the original data files must be available in the samelocation in order to subsequently use load.

The write function makes a copy of the data and saves the copy as a collection of files, which can consume a large amount of disk space. write executes all pending operations on the tall array to calculate the values prior to writing. Once write copies the data, it is independent of the original raw data. Therefore, you can recreate the tall array from the written files even if the original raw data is no longer available.

You can recreate the tall array from the written files by creating a new datastore that points to the location where the files were written. This functionality enables you to create checkpoints or snapshots of tall array data. Creating a checkpoint is a good way to save the results of preprocessing your data, so that the data is in a form that is more efficient to load.

If you have a tall array TA, then you canwrite it to the folder location with the command:

write(location,TA);

Later, to reconstruct TA from the writtenfiles, use the commands:

ds = datastore(location);TA = tall(ds);

Additionally, you can use the write functionto trigger evaluation of a tall array and write the results to disk.This use of write is similar to gather,however, write does not bring any results intomemory.

Supporting Functions

Most core functions work the same way with tall arrays as they do with in-memory arrays. However, in some cases the way that a function works with tall arrays is special or has limitations. You can tell whether a function supports tall arrays, and if it has any limitations, by looking at the bottom of the reference page for the function in the Extended Capabilities section (for an example, see filloutliers).

For a filtered list of all MATLAB functions that support tall arrays, see Function List (Tall Arrays).

Tall arrays also are supported by several toolboxes, enabling you to do things like write machine learning algorithms, deploy standalone apps, and run calculations in parallel or on a cluster. For more information, see Extend Tall Arrays with Other Products.