Yes, that field will show the peak FLOPS of the fastest GPU
in the case of different GPUs of the same vendor.

I added it because the credit formula is essentially
(elapsed time)*(peak FLOPS)*(normalizing factors).
Now we're showing all of these on the web.

-- David

On 07-Jul-2014 2:49 AM, Richard Haselgrove wrote:
David,

I'm a little confused by today's checkin under this title:

http://boinc.berkeley.edu/gitweb/?p=boinc-v2.git;a=commit;h=e437d098243730d71aa531b31860e549bca303f6



To be useful and meaningful on a result page, this line should show the peak 
flop
count of the device the task was actually run on. For a heterogeneous host with
multiple GPUs from the same vendor, the choice of GPU to run an individual task
on is made by the client scheduler at run time.

Yet in sched_send.cpp, line 915, I see

result.flops_estimate = bavp->host_usage.peak_flops;


and from sched_types.h,

struct HOST_USAGE { int proc_type; double gpu_usage; double gpu_ram; double
avg_ncpus; double max_ncpus; double projected_flops; // the scheduler's best
estimate of wu.rsc_fpops_est/elapsed_time. // Taken from host_app_version 
elapsed
time statistics if available, // else on estimate provided by app_plan() double
peak_flops; // stored in result.flops_estimate, and used for credit calculations

If the field you're reporting is populated when you send the job to the user's
host, won't it always report the peak_flops of the 'best' GPU in the host, not
the actual GPU used? That's going to be really confusing.
_______________________________________________ boinc_dev mailing list
[email protected]
http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit
the above URL and (near bottom of page) enter your email address.

_______________________________________________
boinc_dev mailing list
[email protected]
http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev
To unsubscribe, visit the above URL and
(near bottom of page) enter your email address.

Reply via email to