Hi All,

Finally I have got round with an inexpensive Gigabit Ethernet solution
for my Blade 1000.
I have tried first with a Netgear GA311 16bit PCI as recommended in a
previous posting, but they gave me several problems. I tested three
different cards coming from different suppliers.

The blade 1000 complained about "spurious interrupts" with all of them,
filling /var/adm/messages with tons of logs. I tried them on Ultra 5
and they are OK only until the next warm boot (with init 6), than the
box fails to boot reporting panic with errors related to PCI network
(GA311).

Eventually I purchased an Intel pro/1000 (PCI-X), and it works like a
charm but ...

The Blade has 2x900 MHz processors and 5GB RAM + Dual Intel pro/1000
PCI-X

I tried with a basic FTP transfer from/to a dual Pentium III 1.4MHz
running linux (having 16 bit PCI GA311) and I got max 30 MB/s with 60%
CPU taken (25% ftp process, rest is kernel), which means 100% of one
CPU and a bit of the second. The same under the Linux box was in all
taking less than 10% CPU in total.

I tried the same test between the Dual PIII 1.4 and a Dual Opteron
(Linux) and I got 90 MB/s. CPU Used on Opteron around 5%.

I had tried the same on the Blade when it had the GA311 and I got max
30 MB/s (despite the "spurious interrupt" errors).

Why is the Blade so rubbish with the network in comparison with the
Intel and AMD Linux based boxes? Isn't the former supposed to be
optimized for high bandwidth data transfer with fast DRAM, 150 MHz data
bus, etc?

HD

Hi,


haydude wrote:
> Hi All,
>
> Finally I have got round with an inexpensive Gigabit Ethernet solution
> for my Blade 1000.
> I have tried first with a Netgear GA311 16bit PCI as recommended in a
> previous posting, but they gave me several problems. I tested three
> different cards coming from different suppliers.
>
> The blade 1000 complained about "spurious interrupts" with all of them,
> filling /var/adm/messages with tons of logs. I tried them on Ultra 5
> and they are OK only until the next warm boot (with init 6), than the
> box fails to boot reporting panic with errors related to PCI network
> (GA311).
>
> Eventually I purchased an Intel pro/1000 (PCI-X), and it works like a
> charm but ...
>
> The Blade has 2x900 MHz processors and 5GB RAM + Dual Intel pro/1000
> PCI-X
>
> I tried with a basic FTP transfer from/to a dual Pentium III 1.4MHz
> running linux (having 16 bit PCI GA311) and I got max 30 MB/s with 60%
> CPU taken (25% ftp process, rest is kernel), which means 100% of one
> CPU and a bit of the second. The same under the Linux box was in all
> taking less than 10% CPU in total.
>
> I tried the same test between the Dual PIII 1.4 and a Dual Opteron
> (Linux) and I got 90 MB/s. CPU Used on Opteron around 5%.
>
> I had tried the same on the Blade when it had the GA311 and I got max
> 30 MB/s (despite the "spurious interrupt" errors).
>
> Why is the Blade so rubbish with the network in comparison with the
> Intel and AMD Linux based boxes? Isn't the former supposed to be
> optimized for high bandwidth data transfer with fast DRAM, 150 MHz data
> bus, etc?
>
> HD
>
I have not tried any of these adapters, but it depends very much about
the driver itself.

Also MHz is King in this, if you would go back in time or chose "server"
adapters you will find that the adapters is very capable themself.

When GB arrived there was from what I know only two vendors of adapters,
Sun and Alteon, Alteon(Now Nortel) was OEM to Digital/3COM and more.

The Alteon was actually 2xMIPS CPUs inside to offload the CPU due to
poor performance at that time of CPUs.

Then come LevelOne and Intel(82543) which could the PRO 1000 adapter you
have, 82543 was a very broken chip and Intel refused to fix it at that
time so performance was very low compared to Alteons Tigon II chip.

Not much help but MHz is needed or a very good adapter, Alteon Tigon II!


/michael

To further minimize variables, you might want to use say a netperf
TCP_STREAM test between systems. That will eliminate filesystem from
the equations.

When you do the CPU utilization measurements on the Sun system be
certain to read the comments in the relevant files - the "kstat"
files. And the relevant READMEs.

happy benchmarking,

rick jones
--
web2.0 n, the dot.com reunion tour...
these opinions are mine, all mine; HP might not want them anyway...
feel free to post, OR email to rick.jones2 in hp.com but NOT BOTH...

Rick Jones wrote:

> happy benchmarking,

Rather than benchmarking I am mostly interested in actual real life
performance. Within my private network I make extensive use of NFS and
samba. That was the main reason behind the upgrade to gigabit. Disk I/O
performance of the Blade is reasonably good. It is a pity I can't see
that through real life storage shares.

HD