The servers for Folding@home V7 apparently return invalid “work server” IPs to clients requesting Work Units for unsupported GPUs.

The client software prints no explanatory error message, leaves the unsupported GPU’s “compute slot” enabled (taking 1 CPU core out of commission with it), and indefinitely continues to attempt to connect to the obviously-dummy IPs to request Work Units, interleaving bizarre and confusing connection errors in the logs.

This behavior is entirely undocumented, save some passing comments from staff or otherwise-informed posters, indistinguishable at-a-glance from the rest of the ocean of topical speculation, in each of the many threads on this topic.

Before realizing that these dummy IPs are being given out intentionally, I posted the following very clever “workaround”:

sudo ip addr add 192.0.2.1/24 dev `cd /sys/class/net ; ls -d1t e* | head -n 1`
sudo ssh -v `whoami`@localhost -L 192.0.2.1:80:40.114.52.201:80\
 -L 192.0.2.1:8080:40.114.52.201:8080 tail -F /var/lib/fahclient/log.txt

As you can probably deduce, this will forward all TCP connections to 192.0.2.1 on ports 80 and 8080 to the IP 40.114.52.201:80—a legitimate work server.

Sadly/thankfully, these servers seem to return empty responses to clients reporting unsupported GPUs—it would certainly have been embarrassing to retrieve a sequence of GPU Work Units and then be unable to complete them due to missing cl_khr_fp64

12:34:56:ERROR:WU00:FS01:Exception: 10002: Received short response, expected 512 bytes, got 0

Leave a Reply

Your email address will not be published. Required fields are marked *