xCalls: safe I/O in memory transactions
Proceedings of the 4th ACM European conference on Computer systems
Notary: Hardware techniques to enhance signatures
Proceedings of the 41st annual IEEE/ACM International Symposium on Microarchitecture
| Hi-index | 0.00 |
Suboptimal performance of the ISC BIND9 DNS server with multiple threads is a well known problem. This paper explores practical approaches addressing this long-standing issue. First, intensive profiling identifies major bottlenecks occurring due to overheads for thread synchronization. These bottlenecks are then eliminated by giving separate work areas with a large memory pool to threads, introducing faster operations on reference counters, and implementing efficient reader-writer locks. Whereas some of the solutions developed depend on atomic operations specific to hardware architecture, which are less portable, the resulting implementation still supports the same platforms as before through abstract APIs. The improved implementation scales well with up to four processors whether it is operating as an authoritative-only DNS server, with or without dynamic updates, or as a caching DNS server. It also reduces the memory footprint for large DNS databases. Acceptance of this new sever will also have a positive side effect in that BIND9, and its new features such as DNSSEC, should get wider acceptance. The direct result has other ramifications: first, the better performance at the application level reveals a kernel bottleneck in FreeBSD; also, while the results described here are based on our experience with BIND9, the techniques should be applicable to other thread-based applications.