Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

change notes

Browse Source
This commit is contained in:
tqchen 2014-12-04 09:09:29 -08:00
parent 65a1cdf8e5
commit f9d634ce06
1 changed files with 0 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
README.md
View File

@ -8,7 +8,6 @@ Design Note
==== ====
* Rabit is designed for algorithms that replicate same global model across nodes, while each node operating on local parition of data. * Rabit is designed for algorithms that replicate same global model across nodes, while each node operating on local parition of data.
* The global statistics collection is done by using Allreduce * The global statistics collection is done by using Allreduce
* Currently, Rabit is not good at problems where model is distributed across nodes, other abstractions might suits the purpose (for example [parameter server](https://github.com/mli/parameter_server))
Design Goal Design Goal
==== ====
@ -28,8 +27,3 @@ Features
* MPI compatible * MPI compatible
- Codes using rabit interface naturally compiles with existing MPI compiler - Codes using rabit interface naturally compiles with existing MPI compiler
- User can fall back to use MPI Allreduce if they like with no code modification - User can fall back to use MPI Allreduce if they like with no code modification
Persistence of Program
====
Many complicated Machine learning algorithm involves things like temporal memory allocation, result caching.
It is good to have a program that persist over iterations and keeps the resources instead of re-allocate and re-compute the caching every time. Rabit allows the process to persist over all iterations.