Re: [rmcat] draft-sarker-rmcat-eval-test-00, test case 4.4 RMCAT Flow competing with a long TCP Flow

["Eggert, Lars" <[e-mail]>]

Hi,

one additional source of reordering we've seen is caused by the receiver stack itself. At least Linux currently seems to introduce it when the system gets fully loaded (e.g., with 40G NICs) and the scheduler moves queue processing between cores. (This is anecdotal at the moment, we're trying to quantify/qualify it better.)

I'm sure this is a transient problem as far as the kernel source is concerned, but boxes based on the current kernels may live in deployments for quite some time before being updated.

Lars

On 2014-4-30, at 19:02, Fred Baker (fred) <[e-mail]> wrote:

> 
> On Apr 30, 2014, at 4:50 AM, Zaheduzzaman Sarker <[e-mail]> wrote:
> 
>> Are you saying re-ordering is not a phenomena that we can observe in the current Internet? And we should not have test cases to handle re-ordering of packets in RMCAT?
> 
> Packet reordering, loss, and duplication happen. Where and to what degree is a question of the structure of networks that packets pass through and what happens in those networks. The last time I know of someone doing an actual measurement and reporting the results was rather a while ago, but it said that at that time about 5% of arriving traffic was reordered. For systemic reasons, I would imagine this is reduced, and very dependent on what networks and types of networks are crossed.
> 
> Going out on a limb a bit (this is what I suspect, not what I know), I would imagine that today there are three major sources of traffic reordering. One is loss; in radio networks (which include 3g/etc but also include 802.11, 802.15, and 802.16), loss and retransmission are pretty common. It would be interesting to have someone write a paper measuring and analyzing the resultant “best effort” behaviors in those networks. Another is load balancing in the network; while it is common for ECMP routing to send a subset of flows one way and another subset on another, some networks schedule packets on the most available bandwidth or go round robin. That is a recipe for reordering. The other has to do with adaptive bitrate video and services that deliver it by literally duplicating data. In these services, there is a path from the CDN to the consumer that consists of two disjoint paths. The TCP originator is configured with a link layer multicast address for its first hop router, and there are in fact two such routers, one on each of the parallel paths. The packets don’t arrive at the endpoint, but they do arrive at a point just before it, and the first of each pair of packets to arrive is forwarded to the endpoint. You can imagine that the collection point sees frequent duplication and some degree of reordering.
> 
> The single biggest source of reordering, I will postulate, is none of those; it is the result of loss due to TCP’s AIMD behavior. When a packet is dropped in flight due to queue depth or delay, the rest of the packets in that RTT arrive before it, and then its retransmission arrives. Voila! Packet reordering.
> 
> Should TCP handle packet reordering? Yes, of course. Making packet arrival sequence completely random will certainly test that behavior; we could discuss the beneficial effects of running TCP over CSNET in the 1980’s and the resulting hardening of TCP’s algorithms. That doesn’t make that level of reordering a valid simulation of the Internet. A more valid simulation will have one or both of two characteristics - that a packet is dropped, the remaining packets in that RTT pass it, and then it arrives, or that packets frequently arrive disordered but to a degree that doesn’t trigger TCP’s Fast retransmit algorithm.