From: Metod K. <met...@lu...> - 2013-02-25 14:59:58
|
Howdy! I'm quite sure that "dropping" devices will drop any kind of packet when in buffer full state. This is particularly true for layer 2 devices (such as Ethernet switches and/or adapters) who know nothing about IP stuff. If you're lucky this will not be a completely sporadic process if those devices know about QoS, such as TOS(Type Of Service). This behaviour will not show too extermely for TCP due to several factors: * TCP slow start. Transmit rate at the beginning is quite slow for TCP and if sender receives ACKs in timely fashion and no NACKs in between, the TX rate will get up. This is also governed by initial/maximum TCP window size. * as already mentioned in paragraph above, if there's a missing packet, receiving side of a TCP stream will request for retransmission by NACKing the missing packet. This will reduce transmission rate. * probably there are more factors If there's a box with too small buffer size on the link path (leaky wireless device being most suspicious), TCP will handle that through ACK/NACK/window size mechanism and you will not see many retransmissions - most probably less than 10%. BTW, TCP will also handle out-of-order arrival of packets. If using UDP, application will see (and will have to handle) such out-of-order packets. As to the iperf itself: retransmissions are transparent to userland application and iperf receiving side (eg. iperf server) will not see if there are missing TCP packets at all. The only indication of link misbehaviour will be reduced end-to-end throughput. Similar effect will have large link latency - any packet needs to be ACKed and if ACKs arrive late, sending party will not send out more, or in worst case it will already re-transmit non-ACKed packet. The only way to see missing and/or retransmitted TCP packets is via some capturing software (such as wireshark). BR, Metod Hickman, Mark je dne 25/02/13 15:16 napisal-a: > > Thanks for your response. We were coming to the conclusion that iperf must > measure the buffer fill rate on the client side. I am glad that you can > confirm this. On systems with larger buffers we can see the dynamics you > described where iperf can initially fill the buffer rapidly then settles to > the link bandwidth as the test progresses. > > When testing under UDP we are convinced that packets are flushed out of the > buffer without being pulled into the radio. We are having difficulty > determining if a similar buffer behavior is present with TCP traffic. > Though such behavior may be better tolerated due to the TCP retry mechanism > I am concerned that it would at least trigger more retries at the TCP level > than necessary. > > Again, thanks. > > Regards, > > Mark > > *From:*Metod Kozelj [mailto:met...@lu...] > *Sent:* Monday, February 25, 2013 8:43 AM > *To:* Hickman, Mark > *Cc:* ipe...@li... > *Subject:* Re: [Iperf-users] iperf: How does iperf -c xxxxx -1 y -b ??M > calculate the bandwidth at each interval? > > Howdy! > > It seems like nobody answered this question. Or I never got the answer. > Anyhow ... > > Hickman, Mark je dne 15/02/13 22:56 napisal-a: > > We are confused about the inner workings of iperf which raises questions of > the validity of the reports under these conditions. > > 1.Does the iperf client measure the interval bandwidth by the amount of data > it wrote do a buffer per interval or by the amount of data the radio pulled > out of the buffer per interval or the amount of data the radio put on the > air? The first two should be equivalent. > > 2.The other question I should ask but cannot think of. > > > iperf as a perfect userland application does not have any idea about > physical layer connectivity. Hence sending party measures bandwidth of > pushing data into send buffers. It is known to iperf if layers below that > (TCP/UDP; IP; ethernet or any other L2 technology; wire, wireless or any > other L1 technology) drop data. > The above statement does not imply which of the cases you enumerated is > actually the correct. However, if the transmit device behaves (ie does not > drop data due to buffer full), then one an observe typical behaviour: a > surge of UL data with high peak throughput at the beginning and a drop to > real L2 speed afterwards. Hence conclusion: iperf client measures the > interval bandwidth by the amount of data it wrote do a buffer per interval > > If it was the second (iperf client measures the interval bandwidth by the > amount of data the radio pulled out of the buffer per interval), one could > not see the spike right at the beginning of test. > > The visibility of this spike is proportional to the size of transmit buffer > size (wmem in linux) and inversely proportional to the first leg link speed. > > Now to the drops: I have extensive experience with broadband wireless > (WCDMA/HSPA and LTE) devices and most of them are transparent in a sense > that they don't buffer data. A few of them buffer data (and few of them even > act as a kind of router performing NAT etc) and those are more than happy to > drop packets. > When using the former breed one can see the TX speed at the sending side > with no (or seldom) dropped packets while when using the later ones one can > only see the real throughput on the receiving (iperf server) side ... and > there are plenty of dropped packets, amount depends on the ratio between > iperf tx bandwidth versus real link bandwidth. > > In case of high buffering on the way, reports from iperf server tend to be > too late for the client to make note of them. All in all, it's safest to > only rely on reports from the receiving side (server for uplink and client > for downlink). > > BR, > Metod > |