Lecture 12

Geo distributed data centers I am Dr Raji Mishra working as professor in the department of computer science and engineering Indian Institute of Technology partner content of this lecture in this lecture we will Zoo distributed cloud data center interaction of the data centers with the users and the interaction of data center with other data centers in this lecture we will also describe the data center interconnection techniques that is how two data centers they are interconnected such as the techniques such as MLS or Cutting Edge Technologies which the Google data center uses called Google's B4 design and the Microsoft Swan system that we are going to cover in this particular lecture so let us understand the inter data center networking so the virtual use of any powerful web application means that how to communicate with the server is the data center for example this is the client and this is the data center one and this is another data center two they are all connected together with the the internet so any popular web application let us say that [Music] www.google.com requires to communicate with the server and this is the client this is called client and here there are servers so any virtual use of any popular web application means to communicate with the server in a data center for example here it has to communicate how are the connectivity of this service depends upon the internet so internet becomes crucial in the applications application service performance now the data centers also communicate with each other not only the clients communicate with the data center but data center in turn they are also communicating with each other over the internet using the internet for example replicating the client data across multiple data center is one such example so in the cloud scenario the white area connectivity or the internet is the crucial as the data center infrastructure now the data centers which are connected to the internet or the client which is connected to the internet is often called a wide area Network so wide area network is used here or the internet is the crucial for data center infrastructure now the question is why multiple data centers are often needed so the provider such as Google often need an extensive infrastructure so with many location across the wide expense of the globe this is to achieve the better data availability so if one place the facility is down due to some event of disaster then still the data can be available from other location if it is replicated that is why to ensure the data availability it requires the multiple data center is spreaded across the geography second important thing is called load balancing so when multiple facilities are available then they can spread the incoming and outgoing traffic balancing the load across these particular data center over the Internet third important thing is the latency so if present in multiple parts of the globe then it can reach the clients in a different location at a smaller distances thus reduce the latency fourth important thing is local data loss now several Authority might require the companies to store the data from that company in that jurisdiction itself so because that different countries have their different data laws and the local data laws has to now governing how the data can be stored in which part of the country has to abide by that jurisdiction of that country then finally multiple data centers are needed because of hybrid public Cloud operation so hybrid public Cloud operation can handle the average demand for service from the private infrastructure and then offload the peak demand to the public Cloud So when you say a hybrid of a public private Cloud operation then for exceeding the load that is when there is a peak when the load increases the existing capacity of the private infrastructure then it will use the technology technique which is called offloading and it will offload the compute to the public Cloud so it is using private as well as public Cloud so it is a hybrid public and private Cloud operations now let us understand about inter data center traffic now you can see that there are different data centers let us say in this example there are five different data data centers and five data centers in this figure and these data centers are connected with the internet using border router and often or that Network some of them some of these data centers are using PN service connecting different type of data centers so this particular traffic which is being incurring across different data center called inter data center traffic so if you understand this inter Data Center Traffic then we can understand what kind of network whether the network design can be improved so that the traffic can be accommodated so let us understand about this inter Data Center Traffic characteristics and it uses the Yahoo data sets so study is from this case study is from five different Yahoo Data Center and the data set was drawn from 2011 this study which we are going to cover in this part of the paper that is a first look at inter Data Center Traffic characteristics via Yahoo data sets which was published in 2011 is based on anonymized anonymized traces from the bower routers that connect to these particular data centers now here you can see that there are two plots this is the first plot and this is the second plot showing the number of flows between the client and the data center so here you can see that from data center to the client flows is shown here in this particular figure whereas data center to Data Center is shown in the other figure the three lines on each plot for three different data center locations so here it is location one location two and location three here also location one location two and and location three now here if you see the y- AIS on these two plots are quite different so here you can see the number of flows and in this figure number one you can see the maximum is 10 * 10^ 6 number of flows whereas on the right side that is between data center to Data Center the number of flows is maximum of 1.5 into 10^ 6 that is between data center the number of flows are less compared to the number of flows which are more between client to the data center so in terms of number of flows the traffic between the data center is 10 to 20% of the traffic from the data center to the client so it is 20% more traffic between data center to the client as compared to the data center to the data center traffic now the flows between the data center could be long lived here it is shown and can carry more bytes than between the client and the data center that is quite obvious now the question is why these networks are different so they are persistent dedicated hundreds of gigabits per second connectivity between a small set of end points so that is what is shown over here and this is nothing but the internet that is end host to the application end host and wide area network and private wide area network that is data center so as far as the Microsoft is concerned now Microsoft has an expensive resource with an with amortized annual cost of hundreds of millions of dollars is spent in this particular way of interconnection therefore let us see some of the design issues the first one is very traditional wide area network approach that is through MLS multi-protocol label switching so the traditional approach to the traffic Engineering in such Network uses now MLS that is multiprotocol label switching now here the network with different sites can see or defined area connected to each other perhaps over the long distance fiber links so the networks with several different sites spread over defined area then connected to each other perhaps over the long distance fiber links so link State protocol that is ospf Isis that is called link State protocol that used to flood the information about the network topology to all the notes so at the end of such a protocol every node has the map of the network now the available bandwidth information using flooding so so for traffic Engineering also flood also spread the information about the network usage on different links in the network now given that there is already traffic flowing in this network some links will have the spare capacity and some don't so both Isis and OSF have the extension that allow the flooding of the available bandwidth information together with their protocol messages now fulfill the tannel provisioning requests so knowing that the set of networks when the router receives a new flow set of request it will set up the tunnel along the shortest path on which enough capacity is available now it sends protocol messages to the router on the path setting up the tunnel further mpls also supports the notion of priorities thereby if a higher priority flow comes in with the request for a path low priority flows might be displaced now these flows might then use higher latency or higher cross paths through the network Now update the network state that is the Flood information so after the flow is assigned a terminal the routers also update the network state so update the network State and Flood information so when a packet comes into Ingress router the router here the router looks at the packet's header and decides what label that is what tunnel this packet belongs to then it encapsulates this packet with that tunnel's label and send it to the tunnel now ESS router then decapsulates the packet looks at the packet header once again and send it to the destination in this scheme only the Ingress and egress routers read the packets header and every other router on the path just looks at the assigned labels now simple forwarding across along the path so making forwarding along the path is very simple this is the reason why MLS it is called multiprotocol label switching also mpls can run several different protocols as long as Ingress and egress routers understand that protocol and can map onto the label that is why the name is multiprotocol label switching now this MLS suffers from inefficiency the first problem is inefficiency in terms of usage of expensive bandwidth that is what we have seen in the previous slide that Microsoft is incurring huge amount of expense over the public bandwidth use of public bandwidth and here it is shown that the problem is due to the inefficiency in terms of usage of expensive bandwidth typically this network would be provision for the peak traffic so if you see this particular image that if you have the traffic over the time on the Y AIS utilization provision the network for the peak traffic so the network is provisioned for this particular Peak traffic bandwidth now the bean usage of the network might be very small so this is the mean usage so mean usage is let us say the half of what is the peak traffic and if let us say the network is provisioned with the peak traffic so in this case it's 2.17 times smaller than the peak traffic so most of the traffic is actually in that case is the background traffic here you can see this is the background traffic with some latency sensitive traffic as well so this is called non-background traffic which is called latency sensitive traffic so you can provision for the peak of the latency sensitive traffic and then fill the Gap with the background which is not the latency sensitive so unless you differentiate the traffic by the service you cannot do such an optimization and and this is not easy to do so with the MLS because it does not have the global view of what services are running in the network what part of the network are using and as such so also the related point is that regardless of whether they are multiple services or not MLS the routers make the local gritty choices about the scheduling flows so the traffic engineering here is suboptimal for these reasons such networks typically run around 30% utilization to have enough head room for these inefficiencies and this is expensive now second problem is about the in inflexible sharing another big problem of MLS approach is that it only provides the link level fairness so at any link the flows can share the capacity fairly but this the blue flow also shares the capacity with the red flow over here but the blue and the green flow both get the capacity half the red flows because the red flow uses the multiple path so we have l link level fairness but we do not have the network wide fairness here in this case so Network wide fairness is hard to achieve unless you have a global view of the network so therefore let us see the approach which Google uses that is it does not use MLS because of the drawbacks which we have pointed out in different problems bandwidth of sharing and so on so therefore let us understand about The Cutting Edge wide area network traffic engineering that is by Google's B4 system so Google B4 published a paper in ACMC com 2013 that is B4 is an experience with globally deployed software defined word area network and simultaneously Microsoft Swan also was published in ACM ccom 2013 the title of Microsoft Swan is to achieve achieving High utilization with software defined wi area network so let us understand this concept that is V4 Google V4 first and then so it it will leverage the service diversity that is some tolerate delays are there now to get very high bandwidth utilization wide area network because of the natural fluctuation of the time so to Leverage The diversity in the services some Services need a certain amount of bandwidth that is certain moment of time and they are inflexible and some other services can use to kind of fill in wherever the room is left over so therefore a centralized traffic engineering using software defined Network and open Flow was suggested by Google B system so software defined networking approach will gather the information about the state of the network it will make the centralized decision about the flows of the traffic and then push those decision down to the lower levels to actually Implement them but bringing all that information together in one place is relatively complex decision now exact linear programming is too slow now traditionally with optimization techniques like linear programming which is a way to take set of constraints on the required amount of data flow over the parts of the network and come up with an optimal solution but to apply it to the situation where we need to make relatively quick decisions part of the complexity comes from multitude of services the different priorities if we have one service we could run it in flu algorithm and that must would be much faster so it requires something faster if it's not guaranteed to be exactly optimal Dynamic relocation of bandwidth so the demands on the network change over the time so to make the continual decisions about what traffic is the highest priority to move across which Links at a given moment is challenging with the linear programming to make quick decisions so these are all online algorithms but they are not online in the same way as the things inside the data center might be for example Google runs its traffic engineering 500 or more times a day H so it's not as fine grain as the things we might need inside the data center traffic between these facilities is relatively stable it seems so Edge rate limiting the common commonly commonality in the architecture is to implement an enforcement of flow rates so when the traffic enters the network and will do that at the edge rather than at every hop along the path so let us see the B4 that is Google wide area network which was published in 2011 so Google B4 was the first highly visible software defined Network success so it is a private wi area network connecting Google's data center across the planet and it considered 12 different locations of data center which are spread across three different continents this is one continent second continent and third continent and together there are 12 different data centers which are spreaded here in three different continents now it has the number of unique characteristics first is the massive bandwidth requirement deployed to the modest number of sites so the number of sites are only 12 but the bandwidth requirement is the massive here second is the elastic traffic demand that seeks to maximize the average bandwidth and third is full control over the Ed servers and the network which enables rate limiting and demand measuring at the edge what happens at one site is shown now in more detail in one of the data center so Google's B uh B4 View at one side inside one side it has the data center network of the cluster and Border routers So In traditional setting these are connected using uh ebgp to run the routers which would then also interface using ibgp or Isis to other data centers for finer control over the routing this this would be moved to a software router like here that is qua software based router switch that runs on the server so the interface with the open Flow to set up the routing rules on these particular routers so qua will run the routing protocols between the cluster B routers and also other sites and then OpenFlow uses the routing information from guaga and sets up the forwarding rule in the wide area network routers so here you can see all the things are happening now we have a software control and the traffic engineering server which manages what rules exactly are installed so therefore you can see here in this traffic engineering how in the Google B4 system is happening so so there is a traffic engineering server te now it's a centralized server which connects the topology information the available bandwidth information and has the last information about the flow demands across different sites so you can see this information is being collected over here in the traffic engineering server Now traffic engineering server pushes out the flow allocation to different data centers at the data centers the multiple controllers then enforce these flow allocation on these particular centers therefore bgp routing will become like a very big red switch here you can see it also keeps available the bgp forwarding State because each of this particular switch allows them to have multiple forwarding table at the same time they can afford to have bgp forwarding tables in addition to the traffic engineering if the traffic engineering scheme does not work they can discard the routing table and use the bgp forwarding in so bgp routing table serves as a red big switch now design choices the hardware is here in this case is a custom built Hardware it's cheap and quite simple in 2011 there were not many software defined networks switch available in the market so Google build their own cheap with cheap commodity equipments software smarts with the Safeguard so most of these smarts then resides in the software the switches are simple in the software which is the open Flow logic at each site replicated for fall tolerance using pxas the scalability of the system they use hierarchy of the controllers the software solution achieves near 100% utilization and solves the trafficking problem in3 seconds now there are different design choices so for that it has the hierarchy of controller so at the top level there is a global controller which is talking with the sdn uh Gateway so here you can see it will talk to the sdn Gateway the Gateway can be thought of as the super controller that talks to the controllers at all different data center sites so each site itself have the multiple controller here you can see this is inside this particular site a controller you have many controllers are there side B controller side C [Music] controllers so the hierarchy of the controller simplifies the thing from Global Perspective now another important thing is about the aggregation that is flow groups and Link groups so Al the traffic engineering at this global scale is not at the level of mutual flows but flow groups that also helps scaling further each pair of the sites is not connected by just one link there are massive capacity links that are formed from the trunk of several parallel high capacity links all of these are aggregated and exposed to the traffic engineering layer as one logic link it is up to the individual side to the partition traffic Multiplex and demultiplex traffic across the set of links now we talk about the Microsoft Swan design so Microsoft has publicly disclosed the design for optimizing the wide area traffic flow in their wide area network interestingly interesting feature in this design is to to make changes to the traffic flow without causing the congestion so in Microsoft Swan you can think of different ways to do this but we are always going to run some link that may get used by both the Flows at the same time so let us take the look at the design of the swans that there is an approach to making these updates with certain amount of spare capacity so that the conenction can be avoided this approach can take optimization one step further so providing a pretty strong guarantee on lack of congestion even while the network flow of changing here in this particular example you can see that uh blue flow and the green flow so they are using a different different paths or different routers that is it involves the green flow involves R2 R4 and R3 where the blue flow involves R1 R2 and R3 now if the flow is changed that is the green uh that is the blue is changed to this particular path then you can see that these two flows that is between R4 and R3 both the flows fa and FB will be shared on this particular link instead of this automatically what it does is it will shift the flow green flow between R2 and R3 and R1 R4 and R3 will accommodate the flow fa so fa and FB will be now therefore are simultaneously running without any problem of interference so conclusion in this lecture we have discussed the Z distributed cloud data centers that is the interaction of the data centers and also the interaction of the data centers with with the users and among the data centers we have also discussed various data center interconnection techniques such as MLS Google's V4 and Microsoft one thank you [Music] [Music] [Music]