没有多播的世界

3.2.．1.咖啡因序列启动……我们起飞了!今天，在一个寒冷的威斯康辛州的早晨，我从温暖的床上滚起胖乎乎的屁股，向印度的一群工程师举办了一个WebEx研讨会，主题是“为新视频设计你的网络”。当我告诉他们这不是一个多播研讨会时，我想他们一定很惊讶。我也是……视频在我们如何将其部署到我们的网络上方面确实发生了重大变化。回到5年前，我们曾经说过像语音视频和数据的融合。事实上，这意味着声音:是的，数据:是的，视频:嗯，基于今天的网络摄像头和低效率的抄本……嗯，当然，把AVVID拼成两个V比只拼一个要酷得多。 So engineers like me started laying out rendezvous points, learning our *,G and making heads or tails of IGMPv2 over IGMPv3. Multicast was King and life was good. Plus many other apps and routing protocols also use multicasting so it was fat city all around you know. But then a couple major things happened that changed everything... Video turned into a HD and the pictures were amazing. We started watching the Indianapolis Colts put the hurt on other NFL teams and the picture was so clear you could see a linebackers pants pop out when they farted. High quality video production tools once reserved for the deep pockets of Hollywood were now in the hands of regular folks. Videos became a hit. So much so that we integrated video cameras on cell phones and Flip cameras. Folks wanted video at all times. Don't believe me, fire up a video and see how many folks shoulder surf you to see what you are watching. Even without sound, video rules and captures attention. Just like the over used and tired line from Field of Dreams; "Get me a beer and hotdog so I watch these dudes play ball" PCs followed Mac's (insert snarky Mac elitist comment here) and came with web cams built in and now new VOIP phones have video cams built right in. Video is on your network. When it comes to planning your network for the "NEW" video please do not make the same early mistakes I made early on: - Just because your network can handle voice does not mean it is ready for video. - Don't think it's all about multicasting. Multicast play a small role in video today. - Your on video for the enjoyment of the world...all...the...time... ah college.... Here is why; an HD video transfer breaks down to simple math. In HD parameter of 1080p30 basically means that: - I have 1080 lines of horizontal resolution that are combined with 1920 lines of vertical resolution. HDTV 101 right? you can learn that from a Best Buy ad. Let's take it out farther. 1080p30 means that I have 2073600 pixels per screen. The "p" (progressive) means that every time I receive a frame each and every line is refreshed and redrawn. If you see an "i" (interlaced) that means that every other line is refreshed when I receive a frame. The "30" means that I am receiving frames at a rate of 30 per second. That's a lot of refreshing going on per second. In network planning terms putting HD on your network looks like this:2073600像素* 3字节(每个像素持有3字节的信息)* 8位每字节采样* 30帧每秒= 1.5GB的信息!Yowza ! !必须做点什么才能让它在网络上运行。进入h压缩。H.264实际上把这个信息从1.5GB压缩到了5Meg!压缩比是99.8%。有点像真空太空袋之类的东西，用来存放我的旧38特别t恤!非常令人印象深刻的!但问题就从这里开始。以太网本质上是一种损耗介质。一般来说，这没什么大不了的。见鬼，在最好的VOIP网络上，我们可以在1-3%帧损失的情况下工作。此外，随着手机的使用，我们个人对声音不佳的容忍度也提高了。特别是如果你有AT&T。你认为在一个压缩的H.264帧中会丢失多少信息?这肯定是一个值得注意的数量。一帧丢失将导致屏幕像素化(想象中出现的小方块)。声音实际上就像一群训练有素、纪律严明的军队。这是一个很好的行为和可预测的服务。 Oh sure it's time sensitive but the frames are small, constant size (for example g7.11 frame is always going to be 160 bytes plus the overhead and flow at the rate of one every 20mS), low bandwidth requirements, ect... More and more I am seeing voice networks deployed without any QoS. I do not recommend it, but folks are doing it without problems. Makes sense with pipe sizes and high speed ASICs, TCAMs in today's gear. Video on the other hand is like a preacher's kids when they leave home and go to college. They can be all easy going one minute, then nickel beer night hits and well you know the rest of the story... Video has a real "bursty" nature. The amount of information transferred depends upon how much movement is going on. When your screen goes thru a refresh cycle the minimal amount of lines you have to repaint the better. The truth lies in the math so if we are transferring 5M per second @ 30 frames per second that equals out to a frame every 33mS. BUT the amount of info transferred in a video frame depends on the movement so 33mS is an average and can burst way beyond that. Now you know why Telepresence rooms look like as sterile as they do. It has little to do with a unified experience and more to do with a bandwidth conservation methodology. What's important here: - High Availability to the MAX! We can not loose packets. Hang on to those packets like a coupon for free chicken at Popeye's!! Nothing pisses an Executive more then to look like Max Headroom on a video conference. The tolerance is 0.0-.05% loss but it is really 0.0 once you factor in that things only break when management is using it and that happens around performance eval time... For my HD networks, I run ISSU/NSF/SSO on my 4500's and VSS on my 6500 as well as StackWise+ on my edge devices. My routing protocols like EIGRP or OSPF are config'ed for sub-second recovery. I avoid STP if possible and use Multi chassis Etherchannel IF I can. OK, time to look at latency. Our target for latency is similar to that of VOIP 150mS one way. But truthfully, our range is 150-200mS one way. There are three other things to consider when working out your latency budget. Two you can not control on one you can.序列化:就像让你喝醉的表弟离开你家一样，时间是固定的。这是将L2帧转换为L1组脉冲(电/光)所需要的时间。这里需要知道的重要的事情是你的设备上的串行延迟是什么。-传播延迟:这就是从A点到b点的实际延迟时间。基本上，你放了个屁，别人闻到了，然后你就把责任推给了别人。如果你计划延迟，传播将占你预算的95%。这也是你无法控制的，因为这是物理特性。光在光纤中传输15K英里需要120mS。排队:这完全在我们的掌控之中。排队延迟是抖动的主要原因。这就是可靠的QoS规划发挥作用的地方。关于QoS规划/映射的两个权威是RFC 3246，它是一个标准，RFC 4594是一个最佳实践或信息RFC。在HD网络上，你绝对必须积极地使用QoS策略。这是我们唯一需要影响延迟的区域，因为其他两个已经修复。确保你所拥有或使用的开关有螺母来支持你所需要的排队。Cisco用这样的方程列出他们的缓冲区:1p3q3t, 1p7q4t，等等…3q:三个加权轮询队列(分成70/25/5)。3T:三个触发或加权随机早期检测下降阈值。 Man alive, there is so much more to write here. Cisco Press does have a great book on this titled: Cisco Telepresence Fundamentals and some great design stuff is also located at http://www.cisco.com/go/srnd under the video section. Just keep in mind that multicast is NOT the only way to do video now. Truthfully, I am seeing more stream splitting going on now for point to multi-point video and multicast is being pushed down to a protocol handler more and more. When it is used, it is used in Source Specific Mode (SSM). I do not see multicast going away now or ever really as a stare at my USB enabled Pet Rock... Jimmy Ray Purser Trivia File Transfer Protocol Happy Birthday to You; that'll be 30 bucks or the DMCA going to get ya! AOL Time Warner owns the copyright of "Happy birthday to you" until 2030. That's why movies often use different songs for birthday scenes. AOL Time Warner earns over $2 million per year from royalties for the song. Yowza!