diff options
| author | Stefan Metzmacher <metze@samba.org> | 2013-10-14 10:33:57 +0200 |
|---|---|---|
| committer | Stefan Metzmacher <metze@samba.org> | 2013-11-27 14:39:10 +0100 |
| commit | 4244a2686cddcdc754c284df884ae497afa4053a (patch) | |
| tree | 4d3b85a6fbb74b8ab99d15fa7acade2519e59cbf /source3/pam_smbpass/samples/kdc-pdc | |
| parent | 9393e28df59954414313bfae70ffb796d3e332fe (diff) | |
| download | samba-4244a2686cddcdc754c284df884ae497afa4053a.tar.gz samba-4244a2686cddcdc754c284df884ae497afa4053a.tar.xz samba-4244a2686cddcdc754c284df884ae497afa4053a.zip | |
s3:smb2_server: for performance reasons we use tevent_fd and readv/writev directly
Going via tevent_req_create/talloc_free at multiple layer costs
too much cpu cycles per request.
I tested downloading a 16GB (sparse) file with smbclient -b1 -mNT1,
and -mSMB2_02. Using smb2 max read = 64512, which means smb1 and smb2
will use the same read size.
I build with -O3 -g and compared the results with valgrind --tool=callgrind.
With -mNT1 the server uses about 2.000.000.000 cpu cycles.
This patch reduces the userspace cpu cycles for -mSMB2_02
from about ~ 8.000.000.000 down to ~ 4.000.000.000.
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: David Disseldorp <ddiss@samba.org>
Diffstat (limited to 'source3/pam_smbpass/samples/kdc-pdc')
0 files changed, 0 insertions, 0 deletions