1. Scalable Kernel Performance for Internet Servers under Realistic Loads. Gaurav Banga, etc... Western Research Lab : Research Report 1998/06 (Proceedings of the 1998 USENIX Annual Technical Conference) Computer Architecture Lab. CS Dept. KAIST 2000/11/ Kim, Sung-Wan

2. 1/16 Contents Introduction Problems of select() & ufalloc() in event-driven servers Scalable select() & ufalloc() Experimental evaluation Performance of a live system Conclusions

3. 2/16 Introduction Event-driven servers –A single thread manage all connections –Lower context-switching & synchronization overhead faster than a thread-per-connection or pre-forked system –But, perform poorly under real conditions select() & ufalloc() select() –Asynchronous I/O ufalloc() –Allocation of a new file descriptor for a process

4. 3/16 Problems in select() & ufalloc() WAN environments –Larger round-trip time and packet losses than LAN environments –Many open connections select() –select() -> do_scan() -> selscan() -> soo_select() –select_wakeup() -> do_scan() -> selscan() -> soo_select() –soo_select() check to see if the condition is true Linear search for all opened socket ufalloc() –Single bitmap (first lower descriptor number) –Too cost

5. 4/16 Environment Server –AlphaStation 500(400Mhz), 192 MB of main memory –Digital UNIX 4.0B –Squid 1.1.11, NetCache 3.1.2c-OSF Client –AlphaStation 500(333Mhz) –Digital UNIX 3.2C –S-Client Network –100Mbps FDDI Profiling –DCPI

6. 5/16 CPU times in unmodified kernel

7. 6/16 Scalable select() & ufalloc() select() –READY, INTERESTED, HINTS set –sowakeup() Records a hint in the HINTS sets of each of the threads in the referencing processes for which this socket is present in the INTERESTED set of the thread. ufalloc() –2-level bitmap 10 11110011 Level 0 map Level 1 map INTERESTED new = SELECTING U INTERESTED old READY new = C (INTERESTED new ^ (!INTERESTED old U READY old U HINTS)) READY to_user = SELECTING ^ READY new

8. 7/16 Experimental Evaluation - Scalability with respect to connection rate * 750 infinitely slow connections

9. 8/16 Experimental Evaluation - Scalability with respect to connection rate

10. 9/16 Experimental Evaluation - Scalability with respect to connection count

11. 10/16 Performance of a live system Server –A Web proxy system at DEC –AlphaStation 500 (500 MHz), 512 MB of RAM –Running the system for an entire day –Proxy Squid NetCache

12. 11/16 Performance of a live system - NetCache with caching disabled

13. 12/16 Performance of a live system - NetCache with caching disabled

14. 13/16 Performance of a live system - NetCache with caching enabled

15. 14/16 Performance of a live system - NetCache with caching enabled

16. 15/16 Performance of a live system - Squid with caching disabled

17. 16/16 Performance of a live system - Squid with caching disabled

18. 17/16 Performance of a live system - Squid with caching disabled

19. 18/16 Conclusions WAN delays Linear scaling in the select() & ufalloc() –lead to excessive kernel CPU computation Scalable versions –improve the performance of Web servers and proxies

20. 19/16 select(maxfd, &readfds, &writefds, …, …); 1008 for (i = 0; i < maxfd; i++) { 1009 /* Check each open socket for a handler. */ 1010 if (fd_table[i].read_handler) { 1011 if (fd_table[i].stall_until <= squid_curtime) { 1012 nfds++; 1013 FD_SET( i, &readfds); 1014 } 1015 } 1016 if (fd_table[i].write_handler) { 1017 nfds++; 1018 FD_SET(i, &writefds); 1019 } 1020 }