TSP on em makes send of streams very slow
Artem Kuchin
matrix at itlegion.ru
Sat Sep 22 06:07:58 PDT 2007
Kris Kennaway wrote:
> Jack Vogel wrote:
>> On 9/21/07, Artem Kuchin <matrix at itlegion.ru> wrote:
>>> Hello!
>>>
>>> Here is what i have experienced today.
>>>
>>> I just installed 7.0-CURRENT (cvsed and build on 2007/09/20)
>>> on a PRODUCTION web server
>>> (because, IHMO, this current is stable enough and i like
>>> too much :)
>>>
>>> This is intel MB with two built-in em intefaces.
>>>
>>> I sshed to the server.
>>> While i was in plain shell everything was fine, but when i
>>> stared midnight commander i saw how it very slowly draws
>>> scren part by part. It took about 3 monutes to almost
>>> completely draw a screen when i got disconnected. I tied again
>>> - the same. Then i connected via ftp and uploaded 10MB file
>>> at 900KB/sec. When i tried to download it back i got about
>>> 500 *BYTES*/sec and the got disconnected in a couple of minutes.
>>>
>>> Ping was just find, even flood ping from the server on the save
>>> switch with 15000 packets was fine (just one dot on the left).
>>>
>>> I went also crazy already when i desided to compare interface params
>>> with another server with em NICs.
>>>
>>> The dfference is that this is has the following options (by DEFAULT,
>>> i did not turn it on):
>>>
>>> VLAN_HWCSUM,TSO4
>>>
>>> I've read about TSO on 'man ifconfig' and just for kicks decided
>>> to turn it off. VOILA!!! In a seconds send speed was up to 10
>>> MBYTES/sec!
>>>
>>> I have googled about 'em tso slow ' etc.. and found a couple of
>>> seemingly the same problem dated back 2006. Is it supposed to be
>>> solved by now? What IS the problem with TSO?
>>
>> TSO is for some environments, it isn't gonna be useful at 100Mb
>> (which you are), it can be useful at 1Gb but not always, when you
>> get to 10G its
>> a HUGE benefit.
>>
>> Just cuz you can shoot yourself in the foot doesn't mean the gun has
>> a problem :)
>
> So the card can't handle it? Note that the OP says it was enabled
> automatically.
>
> I wouldn't necessarily expect it to give a performance benefit, but it
> shouldn't destroy performance to that extent either. There seems to
> be a real problem to be addressed here.
I have sent this message to freebsdnic at mailbox.intel.com
I took this address from README for em device driver
(/usr/src/sys/dev/em)
But email returned from mailer-daemon because there is no such
email address anymore.
Who is responsible for this driver nowadays?
--
Regards
Artem
More information about the freebsd-current
mailing list