• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Size distributions for memory functions under specific workloads
2
3This folder contains a set of files that are included from `libc/benchmarks/MemorySizeDistributions.cpp`.
4
5Offloading this data to individual files helps
6 - C++ editors (large arrays are usually not well handled by editors),
7 - and allows processing data by other tools to perform analysis or graph rendering.
8
9 ## Format
10
11Most filenames are of the form `{MemoryFunctionName}{WorkloadID}.csv`. They contain a single line of comma separated real values representing the probability that a particular size occurs. e.g.
12 - `"0,1"` indicates that only the size `1` occurs,
13 - `"0.5,0.5"` indicates sizes `0` and `1` occur with the same frequency.
14
15 These files usually contains sizes from `0` to `4096` inclusive. To save on space trailing zeros are discarded.
16
17 ## Workloads
18
19As identified in the [automemcpy](https://research.google/pubs/pub50338/) paper:
20  - `GoogleA` <-> `service 4`
21  - `GoogleB` <-> `database 1`
22  - `GoogleD` <-> `storage`
23  - `GoogleL` <-> `logging`
24  - `GoogleM` <-> `service 2`
25  - `GoogleQ` <-> `database 2`
26  - `GoogleS` <-> `database 3`
27  - `GoogleU` <-> `service 3`
28  - `GoogleW` <-> `service 1`
29
30`Uniform384To4096` is an additional synthetic workload that simply returns a uniform repartition of the sizes from `384` to `4096` inclusive.
31
32## Note
33
34Except for `GoogleD`, all distributions are gathered over one week worth of data.