Old sockfd pool has following defect:
0 statically allocated.
1 use too many fds per sheep, not scalable
2 implemented per thread, can't be shared between threads
3 need resetting at every membership change
The new sockfd cache aims to address these problems yet remain as effecient as
old one:
0 dynamically allocated/deallocated at node granularity.
1 cached fds are multiplexed by all threads.
2 each session (for e.g, forward_write_obj_req) can grab one fd at a time
3 if there isn't any FD available from cache, use normal connect_to() and close()
internally
4 FD are named by IP:PORT uniquely, hence no need of resetting at membership
change
5 the total number of FDs shrinks from (nr_gateway + nr_io) * nr_nodes to nr_nodes
6 just add one more API, totally 3 APIs: sheep_{get,put,del}_fd()
include/sheep.h | 2 +-
lib/net.c | 6 +-
sheep/Makefile.am | 2 +-
sheep/gateway.c | 238 ++++++++++++++++++++++++++++++++++----------------------
sheep/group.c | 9 ++-
sheep/sdnet.c | 66 +---------------
sheep/sheep_priv.h | 14 +++-
sheep/sockfd_cache.c | 407 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
8 files changed, 576 insertions(+), 168 deletions(-)
Liu Yuan (6):
sheep: redesign a new cached sockfd pool
sheep: use new sockfd cache
sheep, sockfd cache: cache more than one FD for each node
sheep: refactor forward_write_obj_req()
sheep: output hex for error code
sheep: try best forward_read_obj_req()
Thanks,
Yuan
|
|