boinc/todo

260 lines
8.4 KiB
Plaintext
Executable File

DON'T ADD ANYTHING TO HERE. USE THE TASKBASE INSTEAD.
File upload handler should check for disk full,
return transient error, alert admins
If there are no active tasks, should start downloading
only the files for a particular result (avoid idle CPU)
Scheduler: don't send results whose deadlines will be missed.
This must take into account:
- pending work on host
- speed of host
- active fraction of host
- network speed of host
- results already in reply
Web site:
show mean, stddev of stats of active hosts
show max potential cobblestones (and percentage achieved)
Web site:
show table of size, MD5 of all executable files
break down by platform
Each RPC should contain a list of projects the host is attached to,
and their resource shares
Add field to user for the above list
Show above info on the web site
Implement coprocessor stuff (see platform.html)
add platform_details field to client state file
pass it in to every app
app can modify it if it wants
send back platform_details field with RPC
store in host table (and in result? make new host if it changes?)
add disk_avail field to host
send in RPC
use this in scheduling
Make sure "update prefs" works even if suspended
In GUI, show "suspended" on tasks and transfers if suspended
Clarify once and for all messages and other logs on Windows;
policy for truncating log files?
-----------------------
BUGS (arranged from high to low priority)
-----------------------
- reset/quit
- Resetting project should delete old project files
-----------------------
HIGH-PRIORITY (should do for beta test)
-----------------------
HTTP stuff
test HTTP redirect mechanism for all types of ops
finish SOCKS implementation, test
use HTTP 1.1
test w/ Apache 2.x
-----------------------
THINGS TO TEST (preferably with test scripts)
-----------------------
server stuff
implement server watchdogs
Add project w/ bad URL or account ID should report error
test this on Win, UNIX
backend stuff
- result reissue
(timeout_check should eventually create new results)
- WU failure: too many errors
- WU failure: too many good results
- credit is granted even if result arrives very late
- shared memory and CPU time measurement, with and without the BOINC API
- ensure cpu time doesn't reset if app is killed rather than quitting
- CPU accounting in the presence of checkpoint/restart
Platform-specific stuff
- timezone on all platforms
- make get_local_ip_addr() work in all cases
per-WU limits
abort result if any file exceeds max_nbytes
max disk
max CPU
max VM size
abort app if excess memory used
Windows screensaver functionality
idle-only behavior without screensaver - test
-----------------------
MEDIUM-PRIORITY (should do before public release)
-----------------------
add an RPC to verify an account ID (returns DB ID for user)
needed for multi-project stats sites
implement a "fetch prefs" command (regular RPC w/o work request)
all RPCs should return a "user-specific project URL"
to be used in GUI (might refer to user page)
in GUI, project name should hyperlink to a project-specified URL
(typically user page for that project)
let user choose language files in installation process
write general language file manipulation functions
use https to secure login pages, do we care about authenticator
being transmitted without encryption from the client?
write docs for project management
how to start/stop server complex
what needs to be backed up and how
account creation: show privacy/usage policies
decide what to do with invalid result files in upload directory
think about sh_fopen related functionality in BOINC client
Implement FIFO mechanism in scheduler for results that can't be sent
user profiles on web (borrow logic from SETI@home)
Devise system for porting applications
password-protected web-based interface for
uploading app versions and adding them to DB
XXX should do this manually since need to sign
Add 2-D waterfall display to Astropulse
Deadline mechanism for results
- use in result dispatching
- use in file uploading (decide what to upload next)
- use in deciding when to make scheduler RPC (done already?)
Testing framework
better mechanisms to model server/client/communication failure
better mechanisms to simulate large load
do client/server on separate hosts?
CPU benchmarking
review CPU benchmarks - do they do what we want?
what to do when tests show hardware problem?
How should we weight factors for credit?
run CPU tests unobtrusively, periodically
check that on/conn/active fracs are maintainted correctly
check that bandwidth is measured correctly
measure disk/mem size on all platforms
get timezone to work
Redundancy checking and validation
test the validation mechanism
make sure credit is granted correctly
make sure average, total credit maintained correctly for user, host
Scheduler
Should dispatch results based on deadline?
test that scheduler estimates WU completion time correctly
test that scheduler sends right amount of work
test that client estimates remaining work correctly,
requests correct # of seconds
test that hi/low water mark system works
test that scheduler sends only feasible WUs
Scheduler RPC
formalize notion of "permanent failure" (e.g. can't download file)
report perm failures to scheduler, record in DB
make sure RPC backoff is done for any perm failure
(in general, should never make back-to-back RPCs to a project)
make sure that client eventually reloads master URL
Application graphics
finish design, implementation, doc, testing
size, frame rate, whether to generate
Work generation
generation of upload signature is very slow
Add batch features to ops web
The Windows installer sometimes leave boinc.# files in the BOINC
directory. This is likely due to the installer not being able to
delete the old boinc.dll file
If a client connects to the scheduling server using default prefs,
use the stored user prefs for determining how much work to send
get preferences works, but is slightly confusing - you have to go to
projects, right click on "get preferences", and then exit/restart boinc
before I get to see my new pretty underwater colors.
"suspend" seems to suspend, but after restart the CPU time jumped up
by a significant amount. This is because Windows 9x uses GetTickCount
for CPU time.
"Retry transfers now" feature, especially for dialup users
-----------------------
LONG-TERM IDEAS AND PROJECTS
-----------------------
CPU benchmarking
This should be done by a pseudo-application
rather than by the core client.
This would eliminate the GUI-starvation problem,
and would make it possible to have architecture-specific
benchmarking programs (e.g. for graphics coprocessor)
or project-specific programs.
investigate binary diff mechanism for updating persistent files
verify support for > 4 GB files everywhere
Local scheduling
more intelligent decision about when/what to work on
- monitor VM situation, run small-footprint programs
even if user active
- monitor network usage, do net xfers if network idle
even if user active
The following would require client to accept connections:
- clients can act as proxy scheduling server
- exiting client can pass work to another client
- client can transfer files to other clients
User/host "reputation"
keep track of % results bad, %results claimed > 2x granted credit
both per-host and per-user.
Make these visible to project, to that user (only)
Storage validation
periodic rehash of persistent files;
compare results between hosts
WU/result sequence mechanism
design/implement/document
Multiple application files
document, test
Versioning
think through issues involved in:
compatibility of core client and scheduling server
compatibility of core client and data server
compatibility of core client and app version
compatibility of core client and client state file?
Need version numbers for protocols/interfaces?
What messages to show user? Project?
Persistent files
test
design/implement test reporting, retrieval mechanisms
(do this using WU/results with null application?)
NET_XFER_SET
review logic; prevent one stream for starving others