Addendum: I tried one ARP1 task. Downloading the input files took two hours with countless retries, but I did not rely on BOINC's built-in retry periods... Executing the task took 21 h on my Haswell Xeon E3 (with all other CPU threads busy with concurrent work, mostly MilkyWay nbody). The seven result files were 94 MB big in total, but they uploaded within a quarter hour without a single interruption. (I had BOINC configured to only 1 transfer/project and 100 kB/s cap on upload speed, which this ARP1 upload fully used.) Either I was lucky and met an unusually quiet point in time at WCG for the upload, or WCG's HTTP troubles only affect downloads, not uploads. Or maybe all work of the current ARP1 batch had already been distributed to hosts and WCG's HTTP errors are over... until the next ARP1 batch.