RE: Iperf-Netperf

[Sahadevan, S. j. d. <sam...@in...>]

After some serious research about Netperf and Iperf it seems like
Netperf is known to flood the receiver side with data as they don't use
any shaping algorithms to control outgoing traffic. On the other hand
Iperf do use shaping algorithm provided by the Linux kernel to control
traffic.

So if my direction of thought is right, is there a way to switch off the
traffic shaping algorithm in Iperf to get bandwidth on par with Netperf?

As for the research I have done on traffic shaping algorithms, it seems
like we need to give commands explicitly to enable traffic shaping
algorithms in the OS before any application (in our case Iperf) might
use it. I haven't done anything like that. 

I will like to know if my direction of thought is right and if so, some
pointers to approach the new direction.


Thanks for the help in advance!
-Samson

-----Original Message-----
From: own...@da...
[mailto:own...@da...] On Behalf Of Sahadevan, Samson
joseph dev
Sent: Thursday, August 24, 2006 4:39 PM
To: ipe...@da...
Subject: RE: Iperf-Netperf

Yeah, I know there is something fishy here... but that is what "sar"
"vmstat" etc., show when I run Iperf.

When I increase the streams from 1 to 3 it goes to 100% busy.

I get a BW of about 90Mbytes/sec for iperf and about 294 Mbytes/sec for
Netperf

-----------------------------------
O-PROFILE REPORT FOR IPERF 1 STREAM I/O SIZE 64
Even O-profile shows that the CPU is 99.9% idle.

samples  %        image name               app name              symbol
name
206322   99.9462  vmlinux                  vmlinux
mwait_idle
43        0.0208  vmlinux                  vmlinux
memmove
7         0.0034  vmlinux                  vmlinux
vgacon_save_screen
3         0.0015  e1000.ko                 e1000
e1000_clean
3         0.0015  e1000.ko                 e1000
e1000_xmit_frame
3         0.0015  oprofiled                oprofiled            (n
symbols)
3         0.0015  vmlinux                  vmlinux
do_page_fault
2        9.7e-04  ext3.ko                  ext3
ext3fs_dirhash
2        9.7e-04  ld-2.3.4.so              ld-2.3.4.so
do_lookup_x
2        9.7e-04  libc-2.3.4.so            libc-2.3.4.so
mbrtowc
2        9.7e-04  vim                      vim                  (no
symbols)
2        9.7e-04  vmlinux                  vmlinux
__do_softirq
2        9.7e-04  vmlinux                  vmlinux
add_softcursor
--------------------------------------
--------------------------------------
O-PROFILE REPORT FOR NETPERF 1 STREAM I/O SIZE 64(top few lines)
(Here the CPU is used 100%)
samples     %        image name               app name        symbol
name
822       7.5447  e1000.ko                 e1000
e1000_xmit_frame
815       7.4805  vmlinux                  vmlinux
tcp_transmit_skb
781       7.1684  vmlinux                  vmlinux
_raw_spin_lock
752       6.9022  vmlinux                  vmlinux           tcp_sendmsg
542       4.9748  vmlinux                  vmlinux
_raw_read_unlock
528       4.8463  vmlinux                  vmlinux           skb_clone
400       3.6714  vmlinux                  vmlinux
ip_queue_xmit
382       3.5062  e1000.ko                 e1000
e1000_clean_tx_irq
---------------------------------------

Any help is greatly appreciated!

-Thanks,
Samson

-----Original Message-----
From: own...@da...
[mailto:own...@da...] On Behalf Of Jonathon Exley
Sent: Thursday, August 24, 2006 3:59 PM
To: ipe...@da...
Subject: RE: Iperf-Netperf

It's funny that you say Iperf is only using 1% of the CPU.
In my experience it uses 100%.

Jonathon. 

________________________________

From: own...@da...
[mailto:own...@da...] On Behalf Of Sahadevan, Samson
joseph dev
Sent: Friday, 25 August 2006 10:40 a.m.
To: ipe...@da...
Subject: Iperf-Netperf



Hi,

To give a better idea of Netperf-Iperf comparison.

The graph explains most of my concerns. (Have the same configurations as
they are taken from the same setup).

Why is Iperf not performing as well as Netperf for lower buffer sizes? 

(Netperf utilizes the whole CPU (100%) to give this performance while if
Iperf keeps almost 99% of CPU idle)

 

Seeing Oprofile data it shows that Netperf spends almost

8% of CPU clock cycles on tcp_sendmsg(), tcp_transmit_skb(),
e1000_xmit_frame() and _raw_spin_lock()

6% of CPU clock cycles on skb_clone() and _raw_read_unlock.  

The remaining are for negligible time duration. 

 



 

 

Can we do some tweaking in the code to increase the throughput of Iperf?

 

Thanks for the help,

Samson 
  

PLEASE NOTE: This email (including any  attachments) is confidential,
and may be
protected by legal privilege and copyright.  If you have received it in
error
please notify us immediately and delete it - it is unlawful to retain,
print,
distribute or reproduce it.  We employ anti-virus software, but cannot
guarantee
emails are virus-free. Views expressed in this email may not be those of
the
originating organisation.