SMB2 performance is worse than SMB1 while iometer 512byte transfer

[Jones]

Hi Jeremy,

2013/9/14 Jeremy Allison <jra at samba.org>
>
>
> talloc_chunk_from_ptr should be inlined, so I'm assuming
> inlined functions shouldn't show up on perf (I may be
> wrong here :-).
>
> How are you compiling smbd. Are you using -O<level> ?
>

You are right! It is my fault!
There is no optimization level is specified when I compiling samba4,
After re-run the configuration phase with CFLAGS="-O -g" and re-compiling,
perf top does not show talloc_chunk_from_ptr anymore.

With samba-3.6, though -O is not specified in configuration phase,
source3/Makefile could see this
CFLAGS= -O -D_REENTRANT -D_POSIX_PTHREAD_SEMANTICS $(FLAGS)
-D_SAMBA_BUILD_=3

With samba-4 series, it is likely -O should be explicit specified,
now samba4 is deployed with python-style waf so cannot found the Makefile
as expected.

Found following links but not get much hint to where add the optimize flags,
is ./configure CFLAGS="-O" sufficient?
https://wiki.samba.org/index.php/Build_Samba
https://wiki.samba.org/index.php/Waf#add_better_optimizations_options



samba-4.2.0pre1-GIT-UNKNOWUN (samba-master on 2013/Sep/12)

Without -O
Protocol   R/W   IOps   %CPU    %us    %sy   %id   %si
-----------------------------------------------------------
SMB2     Read    21410  100%  73.3%  20.8%  0.0%  5.9%
SMB2     Write   20628  100%  73.0%  20.0%  0.0%  7.0%

With -O
Protocol   R/W   IOps   %CPU    %us    %sy   %id   %si
-----------------------------------------------------------
SMB2     Read    27643 100%  67.0%  26.0%  0.0%  7.0%
SMB2     Write   21921 100% 64.0%  30.0%  0.0%  6.0%


"perf top -p <smbd_pid>" SMB2 enabled shows:
     4.05%  libtalloc.so.2.1.0   [.] _talloc_free_internal
     3.11%  [kernel]             [k] ia32_syscall
     2.40%  libc-2.6.1.so        [.] _int_malloc
     1.92%  libc-2.6.1.so        [.] strcmp
     1.45%  libtalloc.so.2.1.0   [.] talloc_alloc_pool
     1.38%  [e1000e]             [k] e1000_xmit_frame
     1.36%  libc-2.6.1.so        [.] _int_free
     1.31%  libc-2.6.1.so        [.] memset
     1.31%  libc-2.6.1.so        [.] __gettimeofday
     1.30%  libtalloc.so.2.1.0   [.] _talloc_get_type_abort
     1.14%  libtalloc.so.2.1.0   [.] _talloc_free
     1.14%  libtalloc.so.2.1.0   [.] _talloc_zero
     1.13%  libtevent.so.0.9.19  [.] _tevent_req_create
     1.06%  libtalloc.so.2.1.0   [.] talloc_get_name
     1.02%  libtalloc.so.2.1.0   [.] talloc_pool
     1.02%  [kernel]             [k] __ticket_spin_lock
     0.90%  libsmbd_base.so      [.] smbd_smb2_request_dispatch
     0.88%  [kernel]             [k] tcp_write_xmit
     0.88%  [kernel]             [k] tcp_sendmsg
     0.86%  libc-2.6.1.so        [.] do_readv
     0.83%  libtalloc.so.2.1.0   [.] __i686.get_pc_thunk.bx
     0.71%  libc-2.6.1.so        [.] memcpy
     0.69%  libc-2.6.1.so        [.] malloc
     0.65%  [e1000e]             [k] e1000_clean_rx_irq
     0.63%  libsmbd_base.so      [.] smbd_smb2_request_process_read
     0.62%  [kernel]             [k] tcp_transmit_skb


Regards,
Jones