& cplSiteName &
Comments
Newest First  |  Oldest First  |  Threaded View        ADD A COMMENT
BigBro
BigBro
12/5/2012 | 5:14:10 PM
re: 10G Ethernet Switches Pass the Test


Sure, the shape tells you a lot about the underlying switch ASIC.


One would expect the latency of a store-and-forward switch to increase with packet size, because the ASIC has to receive the entire packet before it makes its forwarding decision.


On a cut-through switch, you'd expect the latency to be basically flat across packet size, because as soon as the ASIC has received enough of the header, it can make its forwarding decision, and start forwarding the packet (as long as the output port is not busy).


I'm not sure I understand why larger packets would have *less* latency than smaller ones. That's got to be an artifact of the ASIC, or perhaps even the test equipment: once you start getting down into the sub-microsecond range, the test equipment itself becomes a variable in your test that you shouldn't ignore. What MAC and PHY-layer hardware is at that end, for example?

spc_vancem
spc_vancem
12/5/2012 | 5:14:10 PM
re: 10G Ethernet Switches Pass the Test


In the article 97.3 % throughput is said to be low. However, in practical use, is this really a low number? Almost any other traffic types than constant bitrate traffic will result in long queues and full buffers when loads are rising this high. My question is: When is such a high performance actually needed? Any comments anyone?  

Pete Baldwin
Pete Baldwin
12/5/2012 | 5:14:10 PM
re: 10G Ethernet Switches Pass the Test


One thing that surprised me -- and I'd commented to Nick Lippis about this -- was the variety of profiles in the latency testing.  I'm not talking about the actual latency figures, but about the *shapes* of the graphs.


Lippis tested each switch on a variety of packet sizes -- 128-byte packets, 256-byte packets, etc. Some switches had good, low latency for small packet sizes and bad latency for bigger packets. Others were the other way around. Some, IIRC, were consistently flat.


It was interesting to me. It seems to imply that latency effects are rather unpredictable from switch to switch.


As for what causes these different profiles, Lippis was saying a lot of it might be the fingerprint of the chipset being used. Each vendor's different software plays a role, too.

Pete Baldwin
Pete Baldwin
12/5/2012 | 5:14:09 PM
re: 10G Ethernet Switches Pass the Test


> I'm not sure I understand why larger packets would have *less* latency than smaller ones.


Same here. I guess latency is just a tricky beast to wrestle.


I should specify:  Most of the switches either show a mostly flat profile (very good latency for small packets, flatly "less good" for all other sizes) or a sharp, sharp spike for ridiculously large packets (9,216 bytes).  So, some of the visual differences in the graphs come from corner cases. 


But it's true that a couple of boxes showed worse latency with small packets. I found that interesting.

Pete Baldwin
Pete Baldwin
12/5/2012 | 5:14:09 PM
re: 10G Ethernet Switches Pass the Test


You know, that question did occur to me.  But when everybody else is scoring perfect 100s... 97.3 is certainly low by comparison!


You've got a valid question, though, considering not all these datacenter operators will be looking for five 9s kind of performance. Anybody have any real-world experience to apply here?

cross
cross
12/5/2012 | 5:14:07 PM
re: 10G Ethernet Switches Pass the Test


Hi Steinarb,


The average packet size in the Internet is not that small indeed, and keeps growing due to increased video traffic and other bulk applications (see http://www.caida.org/research/.... It is certainly larger than 128 bytes in all cases - the averages measured are somewhere between 150-300 bytes.


"The" worldwide average packet size does not exist, though - it all depends where one measures and which applications dominate in each part of the world - and the situation in data centers is certainly even less uniform. That said, data centers often see an increased fraction of video and storage applications so the average packet size is going up as well according to our findings (but we have no representative proof). Some remote desktop applications (Citrix, for example) and Voice over IP generate small packet sizes around 128 bytes or even less; however, I am not aware of networks where remote desktop or VoIP traffic would drive 10GE ports to full utilization. If a remote desktop application sends bulk screen updates, it uses larger packet sizes as well.


We typically measure the 128-byte single packet size line rate throughput only if a customer explicitly requests it - since the old RFC2544 mentions this measurement as a reference.  The result is of limited value.  Sometimes service providers say they would like to check the chipset's design limitations. In fact, however, today's chipset challenges are more related to bursty traffic of variable packet sizes simultaneously sent, mixed with multicast traffic, coming from many sources and going to many destinations in an imbalanced, meshed traffic pattern.


My alarm bell rather goes off if I see a system reaching 100 % line rate at even the smallest single packet size (64 bytes for IPv4) since it is likely the chipset has been optimized for RFC2544, which does not guarantee its perfection otherwise. We have seen and published such test results in the past.  A good and competitively priced chipset needs to balance throughput requirements of artificial RFC2544 tests with those seen in real-life, complex networks. The art of lab testing is to replicate such real-life scenarios.


Best regards, Carsten (EANTC)


 

tmmarvel
tmmarvel
12/5/2012 | 5:13:57 PM
re: 10G Ethernet Switches Pass the Test


The major throughput and latency problems arise at multi-node network scale, rather than at individual switch scale.

An individual packet switch can be engineered to perform fine on most conditions, but is there a good way for network scale QoS control and throughput optimization in cases of meshed packet streams across multiple packet switches? The operators tend to have to resort to low average network utilization to be able to provide QoS guarantees for services over packet layer shared network 'clouds' serving multiple application/customer contracts.

The major latency, jitter and packet loss problems arise when the volumes of multiple uncoordinated packet streams exceed the capacity of shared physical links, and individual packet switch performance cannot solve these network scale QoS problems inherent to services over multi-client packet layer shared networks.

Which brings up the point of why use packet switched networks as the infra across different service contracts where QoS guarantees are a requirement? WDM/TDM as a mechanism to divide the physical infrastructure between different packet switched contracts certainly eliminates the major network scale congestion control issues. WDM/TDM also supports any packet length mixes up to continued 100% throughput transparently. Within such isolated L1/0 clouds, which can internally provide packet-switched connectivity, the packet layer QoS control will be much more manageable as it can be handled purely from each individual client's edge devices.
 
It would appear that such customer/application level isolation will deal with the bulk of the QoS and throughput problems with packet switched network services. The remaining economic issue is network scale throughput optimization for meshed packet streams, but again, 'better' packet switches do not appear to be the solution there either, as the matter to be focused on is the network scale performance.



stochasticprocess
stochasticprocess
12/5/2012 | 5:13:55 PM
re: 10G Ethernet Switches Pass the Test


In the report, they mention that "For store and forward DUT switches latency is defined in RFC 1242 as the time interval starting when the last bit of the input frame reaches the input port and ending when the first bit of the output frame is seen on the output port."  The latency is not increasing with frame size on store and forward switches because of how the latency is measured (essentially subtracting out the length).  Lippis couldn't have used this measurement method with the cut through switches because this would give you negative latency for jumbo frames (the first bit of output frame would be received before the last bit on the input frame).  Instead you need to do first bit in- first bit out.  Was it the case that Lippis and IXIA measured latency differently for cut through and store and forward switches?

BigBro
BigBro
12/5/2012 | 5:11:57 PM
re: 10G Ethernet Switches Pass the Test


This post explains why latency goes down for larger packets:


http://www.fulcrummicro.com/bl...


"Frame processing time is masked for larger packets. As can be seen in some of the results, the latency gets lower as the size of the packet gets larger. Since the latency clock starts after the last bit arrives, large packets have plenty of time to process the frame header before the first bit is seen on the output port.  With small packets, even after the last bit arrives, the output must wait until frame header processing is complete before the first bit is seen on the output port."



Featured Video
Upcoming Live Events
October 1-2, 2019, New Orleans, Louisiana
October 10, 2019, New York, New York
October 22, 2019, Los Angeles, CA
November 5, 2019, London, England
November 7, 2019, London, UK
November 14, 2019, Maritim Hotel, Berlin
December 3, 2019, New York, New York
December 3-5, 2019, Vienna, Austria
March 16-18, 2020, Embassy Suites, Denver, Colorado
May 18-20, 2020, Irving Convention Center, Dallas, TX
All Upcoming Live Events
Partner Perspectives - content from our sponsors
Edge Computing, the Next Great IT Revolution
By Rajesh Gadiyar, Vice President & CTO, Network & Custom Logic Group, Intel Corp
Innovations in Home Media Terminals for the Upcoming 5G Era
By Tang Wei, Vice President, ZTE Corporation
All Partner Perspectives