Hi David/Eric, On Tuesday 21 July 2009 22:42:47 David Anderson wrote: > > 3) What about signal handling in CUDA apps? How important is it to free > > allocated device memory when the host app is terminated/killed (SIGKILL > > five secs. after SIGTERM) by the BOINC client? I found some indications > > that device memory is implicitly free'd when the host app exists - even > > if you don't call cudaFree(). Can someone confirm this, or is there a > > risk of rendering the GPU device in an unusable state? How about > > page-locked host memory allocated with cudaHostAlloc() or even device > > memory implicitly allocated when creating CUFFT plans...? > > I believe this is handled by the NVIDIA driver; > otherwise crashes of graphics apps would eventually hang the system, > and they don't.
>From the CUDA toolkit release notes: "Sending sigkill (ctrl-c) to an application that is currently running a kernel on the GPU may not result in a clean shutdown of the process as the kernel may continue running for a long time afterwards on the GPU. In such cases, a system restart may be necessary before running further CUDA or graphics applications." This doesn't sound negligible to me. How does SETI handle this for example? Cheers, Oliver _______________________________________________ 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.
