This is good feedback. Unfortunately I don't have time to investigate this fully, but I have a *theory*.
rge exports an empty interrupt blanking function. I suspect that this
could be part of the play. Certainly, it seems ill suited to shifting
into polling mode, at least right now.
Unfortunately, I don't have time to investigate this further, but maybe
someone else can do so.
-- Garrett
Markus Halter wrote:
> Garrett, all
>
> I did some investigations and observed the following. Copying a 1GB file
> using SAMBA and FTP.
>
> >From SAMBA Server to W2K using SMB: 22'
> >From W2K to SAMBA Server using SMB: 2'
>
> >From SAMBA Server to W2K using FTP: 2'
> >From W2K to SAMBA Server using FTP: 2'
>
> >From SAMBA Server to W2K using SMB with SNOOP active: 2'
> >From W2K to SAMBA Server using SMB with SNOOP active: 2'
>
> >From SAMBA Server to W2K using SMB with FTP in parallel: 3'
> >From W2K to SAMBA Server using SMB with FTP in parallel: 3'
>
> Some dTrace measurement shows the following distributions:
>
> [b]Not SNOOPing:[/b]
>
> gtime
> value ------------- Distribution ------------- count
> 256 | 0
> 512 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 1
> 1024 | 0
>
> read
> value ------------- Distribution ------------- count
> 512 | 0
> 1024 |@@@@@@@@@@@@@@ 132
> 2048 |@@@@@@@@@@@@@ 123
> 4096 |@@@@@@@@@@ 98
> 8192 |@@ 17
> 16384 |@ 10
> 32768 | 0
>
> pread64
> value ------------- Distribution ------------- count
> 4096 | 0
> 8192 |@@@@@@@@@@@@@@ 68
> 16384 |@@@ 13
> 32768 |@@@@@@@@@@@@@@ 67
> 65536 |@@@@@@@@@ 41
> 131072 | 1
> 262144 | 0
>
> write
> value ------------- Distribution ------------- count
> 8192 | 0
> 16384 |@@@@@@@@@@@@@@@ 70
> 32768 | 1
> 65536 |@@@@@@ 27
> 131072 |@@@@@@@@@@@@@@@ 72
> 262144 |@@@@ 20
> 524288 | 0
>
> pollsys
> value ------------- Distribution ------------- count
> 1048576 | 0
> 2097152 |@@@@@@@@@@ 48
> 4194304 |@@@@@@@@@@@@@@@@ 75
> 8388608 |@ 6
> 16777216 |@ 5
> 33554432 |@@@ 16
> 67108864 |@@@@@@ 27
> 134217728 |@@@ 12
> 268435456 | 0
>
>
> [b]SNOOPing - rge0 in promiscous mode[/b]
>
> gtime
> value ------------- Distribution ------------- count
> 512 | 0
> 1024 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 5
> 2048 | 0
>
> read
> value ------------- Distribution ------------- count
> 512 | 0
> 1024 |@@@@@@@@@@@@@@ 808
> 2048 |@@@@@@@@@@@@@@@@ 903
> 4096 |@@@@@@@@@ 499
> 8192 |@ 43
> 16384 | 11
> 32768 | 2
> 65536 | 0
>
> pread64
> value ------------- Distribution ------------- count
> 4096 | 0
> 8192 |@@@@@@@@ 232
> 16384 |@@@@@@@@@@@@ 328
> 32768 |@@@@@@@@@@@@@@@@@@@ 538
> 65536 |@ 31
> 131072 | 3
> 262144 | 0
> 524288 | 0
> 1048576 | 1
> 2097152 | 0
>
> write
> value ------------- Distribution ------------- count
> 8192 | 0
> 16384 |@@@@@@ 182
> 32768 |@@@@@@@@ 239
> 65536 | 4
> 131072 |@@@@@@@@@@ 278
> 262144 |@@@@@@@@@@@@@@@ 427
> 524288 | 3
> 1048576 | 0
>
> pollsys
> value ------------- Distribution ------------- count
> 262144 | 0
> 524288 |@@@@@@@@@@@@@@ 407
> 1048576 | 0
> 2097152 |@@@@@@@@@@ 287
> 4194304 |@@@@@@@@@@@@@@@ 422
> 8388608 | 1
> 16777216 | 3
> 33554432 | 4
> 67108864 | 3
> 134217728 | 5
> 268435456 | 0
>
>
> Some how SAMBA is waiting in pollsys() much longer if the performance is
> degraded. Therefor it is not a SAMBA problem. I suspect that somehow the very
> short SMB responses do not trigger an interrupt and the worker thread does
> not handle them fast enough. This might be caused by switching between
> interrupt and polling mode ...
>
> Again this only happens with the rge driver on that Realtec interface. If I
> use gani - which is not a GLDv3 type driver - everything works perfect.
>
> b75 driver does not change anything.
>
> Any ideas ...
>
> Markus
>
>
> This message posted from opensolaris.org
> _______________________________________________
> networking-discuss mailing list
> [email protected]
>
_______________________________________________
networking-discuss mailing list
[email protected]
