Ordinal Technology Corp. - Fast Parallel Sorting
Solutions Performance Product Literature Try Nsort Company Home
Nsort Performance

The Nsort executable and subroutine library have a long history of record breaking sort performance for the MinuteSort benchmark.

Year MinuteSort
Input Reads
Output Writes
2006 40 GB 128 Striped 2587 1221
2004 34 GB 112 Striped 1399 1186

While the MinuteSort contest is now dominated by cluster-based systems, Nsort's records remain the best single-node results (see sortbenchmark.org).

Both of Nsort's records utilized large servers with expensive i/o subsystems. However, current hardware trends have greatly reduced the price of high performance servers and i/o subsystems. Similar or better Nsort performance is possible today at a very modest system costs. But how can good performance be achieved with Nsort?

The performance of the Nsort executable or subroutine library is dependent on many factors including:

  • key types, decimal key type is slowest
  • average record size, bigger record size means better mb/sec
  • processor speed and number of processor cores

However the two most prominent factors that limit Nsort performance are:

  • Nsort memory size
  • disk configuration (when considering the elapsed time to read the input data from disk, sort it, and write the output to disk)

To demonstrate the importance of memory size and disk configuration to Nsort performance, tests were run using an input data set of 100,000,000 100-byte records (10,000,000,000 bytes) as generated by the gensort program used with the standard sort benchmarks.

The system configuration was as follows:

  • Windows Server 2008 Datacenter
  • Intel Core i7-980X Processor (6 cores, 12 hyperthreads)
  • LSI SAS 9207-8i HBA
  • Seagate Barracuda 1TB ST1000DM003 Hard Drives
  • Samsung 840 PRO 256GB Solid State Drives

Various drive configurations were tried using either 1, 2 or 4 hard drives or solid state drives for the sort input and output filse, and temporary files. (A separate system drive was used whose size and type was irrelevant to these test results.) All the sort data drives were connected via the LSI HBA.

These tests do not reflect the highest possible performance with Nsort (more disks --> more better). On the other hand, these tests are optimistic for the number of disk drives used in several ways:

  • Newly formatted disk volumes were used. Real world disk volumes with fragmented free space would yield worse performance.
  • These results were "tuned" in that various file transfer sizes were tried with only the best result presented here. Not all Nsort users try out multiple file transfer sizes.
  • These April 2013 tests used hard drives and solid state drives with excellent performance, not all drives have have comparable performance.

The logical drive type (i.e. number of physical drives used and whether they were single, striped or mirrored) used for the sort input and output files is listed for each test in the first column. If an input and output file transfer size was specified to Nsort, it is listed in the next column. Otherwise this field is left blank. Any drive striping or mirroring was done with the standard Windows Disk Management tool.

The logical drive type for the temporary file is listed in the next column along with any specified transfer size. Some of the tests were done using approximately 11.5 GB of Nsort memory. In these cases the no temporary file was necessary and the temp drive is listed as None. In all other cases, Nsort was directed to use 2 GB of memory thereby forcing it to use the specified temporary drive(s).

The resulting input file read speed, temp file write speed, temp file read speed, output file write speed, and the elapsed seconds of the sort from the Nsort statistics output are all given. The last column contains a link that may be selected to get the Nsort command line and statistics output for the test.

Temp Drive(s) Trans
Input Reads
Temp Writes
Temp Reads
Output Writes
Hard Drives - Seagate Barracuda 1TB ST1000DM003
1 Single None, 1 Pass 179 - - 188 109 Get
1 Single 16m Same as I/O 4m 89 88 85 85 233 Get
1 Single 16m 1 Single 16m 178 180 190 190 113 Get
2 Striped None, 1 Pass 337 - - 371 57 Get
2 Striped 16m Same as I/O 4m 162 160 125 127 143 Get
2 Striped 16m 2 Striped 4m 307 308 214 217 80 Get
2 Striped 16m 2 Single 16m 320 312 308 312 67 Get
2 Mirrored 8m None, 1 Pass 191 - - 188 106 Get
2 Mirrored 16m Same as I/O 4m 92 91 110 111 203 Get
2 Mirrored 16m 2 Mirrored 4m 172 170 191 190 114 Get
4 Striped 16m None, 1 Pass 650 - - 650 31 Get
4 Striped 16m Same as I/O 4m 232 224 195 198 96 Get
Solid State Drives - Samsung 840 PRO 256GB
1 Single 8m None, 1 Pass 563 - - 532 37 Get
1 Single 8m Same as I/O 16m 257 253 273 275 77 Get
1 Single 8m 1 Single 4m 417 411 533 532 44 Get
2 Striped 16m None, 1 Pass 1073 - - 998 20 Get
2 Striped 8m Same as I/O 4m 347 338 451 451 52 Get
2 Striped 8m 2 Striped 16m 679 668 977 978 26 Get
2 Striped 8m 2 Single 4m 961 963 980 979 21 Get
2 Mirrored 16m None, 1 Pass 563 - - 532 37 Get
2 Mirrored 16m Same as I/O 4m 219 216 340 338 77 Get
2 Mirrored 16m 2 Mirrored 4m 383 377 539 532 46 Get
4 Striped 16m None, 1 Pass 2079 - - 1953 10 Get
4 Striped 16m Same as I/O 4m 470 456 875 865 34 Get

Several guidelines can be drawn from these tests.

  • Using enough memory to avoid temporary file usage yields the best results.
  • Striping disks together increases performance, although not quite linearly.
  • Disk mirroring does not hurt performance.
  • If you have multiple disks available, how can they be configured to maximize Nsort performance?
    • If you can give Nsort enough memory so that no temp disk is necessary, then stripe the disks together to form a volume for your sort input and output files.
    • If you can't give Nsort enough memory to avoid temp file usage, the following options are available in decreasing order of performance:
      • Stripe together half of your available disks to form the volume for input and output files. Use the other half of your disks as single disk volumes for temp files.
      • Stripe together half of your available disks to form the volume for input and output files. Stripe together the other half of your disks for use as a temp volume.
      • Stripe together all your disks and use that volume for input, output and temp sort usage. While this yields the worst relative performance, there may be reasons other than Nsort performance to pick this option, such as the most flexible usage of overall disk storage space.