Changes between Version 80 and Version 81 of DevProjects

Timestamp:

Apr 7, 2013, 7:35:38 PM (12 years ago)

Author:

davea

Comment:

--

DevProjects

@@ -45 +45 @@
 Suggest new items in the boinc_dev email list.)'''
 
+=== High priority ===
+
+ 1. Change the GPU framework.
+  Currently BOINC requires that all GPUs of a given vendor (NVIDIA, ATI, Intel) be similar,
+  and it treats them as a single pool
+  (i.e. jobs are not associated with a particular GPU instance).
+  This model has a number of drawbacks on machines with multiple different GPUs.
+  Change the model so that each GPU is treated separately.
+  This will require major changes to the client, scheduler, and RPC protocol.
+ 1. The client's job scheduler has several O(N^2) algorithms,
+  where N is the number of jobs queued on the client.
+  Change these to Nlog(N).
+ 1. The scheduler's logic for selecting app versions is clumsy.
+  Replace it with logic that, at the start, selects a version for each (app, resource type)
+  and stores these in an array.
+
 === Web features ===
 
@@ -58 +74 @@
 
  1. [wiki:PrefsRemodel Remodel the preferences system]
- 1. Combine user page and profile. #601
  1. Change default Q&A page to refer BOINC-specific questions to BOINC web site. #603
  1. Convert team name HTML and team description to BBcode. #604
@@ -78 +93 @@
  1. Investigate the crlibm library for generating identical results across processors
   (or at least reducing the number of cases for HR).
- 1. Java support: core client checks for the existence of JVM, reports version to scheduler.
-   Write Java wrapper (runs JVM, gives it jar files).
-   Note: Sztaki has already done some part of this.
- 1. Same, .NET
 
 === Runtime system ===
- 1. Implement a mechanism where the client can tell apps to checkpoint.
+ 1. Implement a mechanism where the client can ask apps to checkpoint.
  1. Use modern process-grouping mechanisms (e.g. Unix process groups).
 
@@ -91 +102 @@
  1. Keep track of the statistics of how long it takes to upload files, and to report results.
    Use that info to improve compute deadlines (e.g., subtract the 2 sigma point for both).
- 1. Core client monitors total CPU usage, limits its own CPU usage so that total is < 100%.
+ 1. Client monitors total CPU usage, limits its own CPU usage so that total is < 100%.
  1. Monitor working-set size of non-BOINC apps; don't allow total to exceed physical RAM.
    This could eliminate the need for memory-usage preferences.
@@ -128 +139 @@
 === Server and Back End ===
 
- 1. Modify the scheduler's app version selection so that instead of using
-    the fastest available version,
-    it uses the fastest one for the resource R for which R.req_secs/R.ninstances is greatest.
  1. Add support for parallel applications:
    try to schedule all the jobs in a batch more or less simultaneously.