I just installed an HSC 100BaseT network card (2 ports) in a K200 server and
am getting slower than expected transfer rates.

Downloading via FTP from the server I get 4 to around 5 MB/s. I would
normally expect transfer rates of around 8 MB/s with Fast Ethernet.

What could be causing this?

Also NFS is very slow and is causing a lot of read errors.

David Travers <david@nospam.travers.force9.co.uk> wrote:
> I just installed an HSC 100BaseT network card (2 ports) in a K200
> server and am getting slower than expected transfer rates.

> Downloading via FTP from the server I get 4 to around 5 MB/s. I
> would normally expect transfer rates of around 8 MB/s with Fast
> Ethernet.

> What could be causing this?

Given that you are running 10.20 on a K200, my first guess would be
that you are saturating one or more of the CPUs.  While FE is 10X the
performance of "Classic E" it does _nothing_ to make data transfer
easier on the host - it takes just as many CPU cycles to push a KB
across the wire with FE as it did with CE.  So, if you look at your
CPU utilization with Classic Ethernet, you can guesstimate what the
_best_ boost you might see with FE.

The same thing holds true for Gigabit versus FE - certainly to a first
approximation - the _specifics_ of the _implementations_ of GbE NICs
(programming models, CKO etc) can provide an easing of load on the
host, but there is still nothing fundamental in Gigabit Ethernet that
makes life any easier on the host.

My second guess would be that you are encountering some disc/fs
transfer rate limitation.  Compare your FTP perf with that of dd.

You might also want to check what a netperf TCP_STREAM test with 56K
socket buffers can achieve - that will be (iirc FTP used 56KB even
back on 10.20 but I could be wrong) the same socket buffer size as
FTP, but without the filesystem stuff.

> Also NFS is very slow and is causing a lot of read errors.

Check your lanadmin - or was it still called landiag back on 10.20 -
statistics and see if there are any outbound errors.  IIRC the HSC
100BT NICs and their drivers back then tried to set the NIC FIFO
transmit threshold lower to eek-out some better latency - however, the
memory bus could often lose the race and one could get FIFO underruns.
The fix would be to set the transmit threshold to 1500 bytes (aka
store and forward).

-- 
denial, anger, bargaining, depression, acceptance, rebirth...
                                     where do you want to be today?
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to post, OR email to raj in cup.hp.com  but NOT BOTH...

"David Travers" <david@NOSPAM.travers.force9.co.uk> schrieb im =
Newsbeitrag news:IO5Ac.558$qt1.295@newsfe3-win.server.ntli.net...
> I just installed an HSC 100BaseT network card (2 ports) in a K200 =
server and
> am getting slower than expected transfer rates.
>=20
> Downloading via FTP from the server I get 4 to around 5 MB/s. I would
> normally expect transfer rates of around 8 MB/s with Fast Ethernet.
>=20
> What could be causing this?

In addition to what Rick Jones already told you, I would have a look at =
the configuration of full/halfduplex of the machines which you try to =
connect. (You can lanadmin to check the half/fullduplex status of the =
machines)
I have seen ugly things happening in this respect.

HTH

Martin

We have Linux servers that can easily achieve 8/9 MB/s over Fast Ethernet
and almost 30MB/s over Gigabit Ethernet (however these are very fast Xeon
servers with Ultra SCSI 320 hard disks)

I tweaked one of the settings using nettune and now get just over 5.5MB/s,
which is an improvement but still short of the mark.

I'm guessing the main bottle neck is the hard disk and the CPU's (which are
running at a pathetically slow speed of 100Mhz). By the way the system has 4
100Mhz CPU's

The switch (it's a managed switch) reports that the connection is 100MB/s
and full duplex.

Before the upgrade, with normal Ethernet and we were getting only 900Kb/s.
The main reason we have the new card is that we mirror the server's database
(9GB) and speed was the main importance. It was taking 3 hours to perform
the mirror, now it only takes 20 mins.

It's an old system but very important to the company due the complex expert
system it's running.

"Rick Jones" <foo@bar.baz.invalid> wrote in message
news:XW5Ac.4085$342.3170@news.cpqcorp.net...
> David Travers <david@nospam.travers.force9.co.uk> wrote:
> > I just installed an HSC 100BaseT network card (2 ports) in a K200
> > server and am getting slower than expected transfer rates.
>
> > Downloading via FTP from the server I get 4 to around 5 MB/s. I
> > would normally expect transfer rates of around 8 MB/s with Fast
> > Ethernet.
>
> > What could be causing this?
>
> Given that you are running 10.20 on a K200, my first guess would be
> that you are saturating one or more of the CPUs.  While FE is 10X the
> performance of "Classic E" it does _nothing_ to make data transfer
> easier on the host - it takes just as many CPU cycles to push a KB
> across the wire with FE as it did with CE.  So, if you look at your
> CPU utilization with Classic Ethernet, you can guesstimate what the
> _best_ boost you might see with FE.
>
> The same thing holds true for Gigabit versus FE - certainly to a first
> approximation - the _specifics_ of the _implementations_ of GbE NICs
> (programming models, CKO etc) can provide an easing of load on the
> host, but there is still nothing fundamental in Gigabit Ethernet that
> makes life any easier on the host.
>
> My second guess would be that you are encountering some disc/fs
> transfer rate limitation.  Compare your FTP perf with that of dd.
>
> You might also want to check what a netperf TCP_STREAM test with 56K
> socket buffers can achieve - that will be (iirc FTP used 56KB even
> back on 10.20 but I could be wrong) the same socket buffer size as
> FTP, but without the filesystem stuff.
>
> > Also NFS is very slow and is causing a lot of read errors.
>
> Check your lanadmin - or was it still called landiag back on 10.20 -
> statistics and see if there are any outbound errors.  IIRC the HSC
> 100BT NICs and their drivers back then tried to set the NIC FIFO
> transmit threshold lower to eek-out some better latency - however, the
> memory bus could often lose the race and one could get FIFO underruns.
> The fix would be to set the transmit threshold to 1500 bytes (aka
> store and forward).
>
> -- 
> denial, anger, bargaining, depression, acceptance, rebirth...
>                                      where do you want to be today?
> these opinions are mine, all mine; HP might not want them anyway... :)
> feel free to post, OR email to raj in cup.hp.com  but NOT BOTH...

David Travers wrote:
> We have Linux servers that can easily achieve 8/9 MB/s over Fast Ethernet
> and almost 30MB/s over Gigabit Ethernet (however these are very fast Xeon
> servers with Ultra SCSI 320 hard disks)
> 
> I tweaked one of the settings using nettune and now get just over 5.5MB/s,
> which is an improvement but still short of the mark.
> 
> I'm guessing the main bottle neck is the hard disk and the CPU's (which are
> running at a pathetically slow speed of 100Mhz). By the way the system has 4
> 100Mhz CPU's
> 
> The switch (it's a managed switch) reports that the connection is 100MB/s
> and full duplex.
> 
> Before the upgrade, with normal Ethernet and we were getting only 900Kb/s.
> The main reason we have the new card is that we mirror the server's database
> (9GB) and speed was the main importance. It was taking 3 hours to perform
> the mirror, now it only takes 20 mins.
> 
> It's an old system but very important to the company due the complex expert
> system it's running.
> 
> "Rick Jones" <foo@bar.baz.invalid> wrote in message
> news:XW5Ac.4085$342.3170@news.cpqcorp.net...
> 
>>David Travers <david@nospam.travers.force9.co.uk> wrote:
>>
>>>I just installed an HSC 100BaseT network card (2 ports) in a K200
>>>server and am getting slower than expected transfer rates.
>>
>>>Downloading via FTP from the server I get 4 to around 5 MB/s. I
>>>would normally expect transfer rates of around 8 MB/s with Fast
>>>Ethernet.
>>
>>>What could be causing this?
>>
>>Given that you are running 10.20 on a K200, my first guess would be
>>that you are saturating one or more of the CPUs.  While FE is 10X the
>>performance of "Classic E" it does _nothing_ to make data transfer
>>easier on the host - it takes just as many CPU cycles to push a KB
>>across the wire with FE as it did with CE.  So, if you look at your
>>CPU utilization with Classic Ethernet, you can guesstimate what the
>>_best_ boost you might see with FE.
>>
>>The same thing holds true for Gigabit versus FE - certainly to a first
>>approximation - the _specifics_ of the _implementations_ of GbE NICs
>>(programming models, CKO etc) can provide an easing of load on the
>>host, but there is still nothing fundamental in Gigabit Ethernet that
>>makes life any easier on the host.
>>
>>My second guess would be that you are encountering some disc/fs
>>transfer rate limitation.  Compare your FTP perf with that of dd.
>>
>>You might also want to check what a netperf TCP_STREAM test with 56K
>>socket buffers can achieve - that will be (iirc FTP used 56KB even
>>back on 10.20 but I could be wrong) the same socket buffer size as
>>FTP, but without the filesystem stuff.
>>
>>
>>>Also NFS is very slow and is causing a lot of read errors.
>>
>>Check your lanadmin - or was it still called landiag back on 10.20 -
>>statistics and see if there are any outbound errors.  IIRC the HSC
>>100BT NICs and their drivers back then tried to set the NIC FIFO
>>transmit threshold lower to eek-out some better latency - however, the
>>memory bus could often lose the race and one could get FIFO underruns.
>>The fix would be to set the transmit threshold to 1500 bytes (aka
>>store and forward).
>>
>>-- 
>>denial, anger, bargaining, depression, acceptance, rebirth...
>>                                     where do you want to be today?
>>these opinions are mine, all mine; HP might not want them anyway... :)
>>feel free to post, OR email to raj in cup.hp.com  but NOT BOTH...
> 
> 
> 
K200 is a PA1.1 system, as such they do not take advantage of the PA2.0 
arch (and the software that is the most recent). Regardless of how much 
you tweak it will not even get close, find a k4xx|5xx box to swap in 
place of it, it will be well worth it. Lots of good used cheap boxes 
available.

David Travers <david@nospam.travers.force9.co.uk> wrote:
> We have Linux servers that can easily achieve 8/9 MB/s over Fast
> Ethernet and almost 30MB/s over Gigabit Ethernet (however these are
> very fast Xeon servers with Ultra SCSI 320 hard disks)

> I tweaked one of the settings using nettune and now get just over
> 5.5MB/s, which is an improvement but still short of the mark.

Which setting did you tweak?

> I'm guessing the main bottle neck is the hard disk and the CPU's
> (which are running at a pathetically slow speed of 100Mhz). By the
> way the system has 4 100Mhz CPU's

That it has four CPU's doesn't really help for a single TCP
connection.  A single TCP connection will not take advantage of (much)
more than one CPU's worth of processing horsepower.  More than one CPU
could help for mulitple concurrent TCP connections.  You do still need
to check and make sure that one of the CPU's isn't pegged running
interrupts.

You might try some netperf TCP_STREAM tests to see what you can get
absent the discs.

> The switch (it's a managed switch) reports that the connection is
> 100MB/s and full duplex.

Does the switch report any errors (I'm still thinking about the
transmit threshold stuff, might want to check the landiag stuff).

> Before the upgrade, with normal Ethernet and we were getting only
> 900Kb/s. 

And what was the peak CPU utilization on any of the CPUs?  Divide that
into 100% and that is the _maximum_ speed-up you should expect.  So,
if one CPU was up to 20%, you shouldn't expect to get more than 5X
your classic Ethernet throughput.

> The main reason we have the new card is that we mirror the server's
> database (9GB) and speed was the main importance. It was taking 3
> hours to perform the mirror, now it only takes 20 mins.

> It's an old system but very important to the company due the complex
> expert system it's running.

Is the software locked to 10.20?  If you need more performance out of
the system and it is that important, you may need to look into a CPU
upgrade, or simply migrating everything to say a K580.

rick jones
-- 
oxymoron n, commuter in a gas-guzzling luxury SUV with an American flag
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to post, OR email to raj in cup.hp.com  but NOT BOTH...