FastSort

Short Description

FastSort sorts input records using a sort key. FastSort is faster than ExtSort, but requires more system resources and sorting is not stable.

Ports

Metadata

Metadata can be propagated through this component. All output metadata must be same.

FastSort Attributes

[1] Estimated record count is a helper attribute which is used for calculating (rather unnatural) Run size automatically as approximately Estimated record count to the power 0.66. If Run size is set explicitly, Estimated record count is ignored.
[2] These attributes affect automatic guess of Run size. Generally, the following formula must be true:
Number of read buffers * Run size * Average record size < Maximum memory

Details

FastSort is a high performance sort component reaching the optimal efficiency when enough system resources are available. FastSort can be up to 2.5 times faster than ExtSort, but consumes significantly more memory and temporary disk space.

The component takes input records and sorts them using a sorting key - a single field or a set of fields. You can specify sorting order for each field in the key separately. The sorted output is sent to all connected ports.

Satisfactory results can be obtained with default settings (just the sorting key must be specified). However, to achieve the best performance, a number of parameters is available for tweaking.

Capt. Eddie to ground control:

FastSort does not preserve order of records with equal key value (sorting algorithm is not stable). If stability is required, please use ExtSort instead.

Sorting Null Values

In ascending order, null values are sorted before strings, numbers, booleans or dates. If you sort data in descending order, null values are sorted after strings, numbers, booleans or dates.
Remember that FastSort processes records in which the same fields of the Sort key attribute have null values as if these nulls were equal.

FastSort Tweaking

Basically, you do not need to set any of these attributes; however, sometimes you can increase performance by setting them. You may have limited memory or you need to sort a large number of records, or these records are too big. In similar cases, you can adjust FastSort to your needs.

Hint!

The memory requirements of the component can be estimated as follows:
heap memory = Number of read buffers × Run size × estimated record size
direct memory = Max open files × Tape buffer size

Run Size
The core attribute for FastSort. Determines how many records form a "run" (i.e. a group of sorted records in temp files). The lower Run size, the more temp files get created, less memory is used and greater speed is achieved. On the other hand, higher values might cause memory issues. There is no rule of thumb as to whether Run size should be high or low to get the best performance. Generally, the more records you are about to sort, the bigger Run size you might want. The rough formula for Run size is Estimated record count^0.66. Note that memory consumption multiplies with Number of read buffers, which in turn grows with Number of sorting threads. So, higher Run sizes result in much higher memory footprints.

Max Open Files
FastSort uses relatively large numbers of temporary files during its operation. By default, the number of temporary files is limited to 1,000. For production systems, it is recommended to set the limit as high as possible, because there is no speed sacrifice, see Performance Bottlenecks. On the other hand, you can lower the limit even further to prevent hitting the user quota or other OS-specific limits and runtime limitations. The following table should give you a better idea:

Note that numbers in the table above are not exact and might be different on your system.

Number of Sorting Threads
Tells FastSort how many runs (chunks) should be sorted at a time in parallel. By default, it is automatically set to 1 or 2 based on the number of CPU cores in your system. Overriding this value makes sense if your system has lots of CPU cores and your disk performance can handle working with so many parallel data streams.

Number of Read Buffers
This setting corresponds tightly to Number of sorting threads - must be equal to or greater than Number of sorting threads. The higher the number of read buffers, the lower chance the workers will block each other. Defaults to Number of sorting threads + 2.

Compress Temporary Files
Along with Temporary files charset, this option lets you reduce the space required for temporary files. Note that compression can save a lot of space but decreases the performance by up to 30%.

Tape Buffer Size
Size (in bytes) of a file output buffer. The default value is 8kB. Decreasing this value might avoid memory exhaustion for large numbers of runs (e.g. when Run size is very small compared to the total number of records). However, the impact of this setting is quite small.

Best Practices

Make sure you have dedicated enough memory to your Java Virtual Machine (JVM). Having plenty of memory available, FastSort is capable of doing astonishing job. Remember that the default JVM heap space (64MB) can cause FastSort to crash. For best results, try to increase the memory value up to 2GB (if possible, while still leaving enough memory for the operating system). How to set the JVM is described in Runtime configuration.

Performance Bottlenecks

• Sorting big records (long string fields, tens or hundreds of fields, etc.):FastSort is greedy for both memory and CPU cores. If the system does not have enough of either, FastSort can easily crash with the out-of-memory error. In such a case, use the ExtSort component instead.

• Utilizing more than 2 CPU cores: Unless you are using very fast disk drives, overriding the default value of Number of sorting threads to more than 2 threads does not necessarily help. It can even slightly slow the process down, as extra memory is loaded for each additional thread.

• Coping with quotas and other runtime limitations: In complex graphs with several parallel sorts, even with other graph components also having huge number of open files, the Too many open files error and graph execution failure may occur. There are two possible solutions to this issue:
  a. increase the limit (quota) This option is recommended for production systems since there is no speed sacrifice. Typically, setting limit to higher number on Unix systems.
  b. force FastSort to keep the number of temporary files below some limit For regular users on large servers, increasing the quota is not an option. Thus, Max open files must be set to a reasonable value. FastSort then performs intermediate merges of temporary files to keep their number below the limit. However, setting Max open files to values, for which such merges are inevitable, often produces significant performance drop. So keep it at the highest possible value. If you are forced to limit FastSort to less than a hundred temporary files, even for large datasets, consider using ExtSort instead which is designed for performance with a limited number of tapes.

See also

ExtSort
Common Properties of Components
Specific attribute types
Common Properties of Transformers