Thanks for your answers.
OK I will definitively use Zookeeper rather file-based synchronization and let you know.
For information, the syncharea folder during our crawl was not accessed by any other process. The server is dedicated to MCF. The OS is Debian 8 and the files are on standard Linux filesystem (ext3). We did not increase the max open files in this server (only on the Solr servers), it is a good thing to investigate, thanks.
Regardless of the change for ZK, is it possible to change this behavior in MCF by automatically stopping the job for example when this exception occurs ?