IPv6 has expanded addressing capabilities, as it uses 128-bit addresses instead of 32-bit addresses. In addition to unicast and multicast, IPv6 has also anycast, which allows a message to be delivered to any host of a set of hosts.
The header has a fixed size of 40 bytes, which simplifies hardware implementations.
The header also has the traffic class (8 bits) and the flow label (20 bits) fields.
IPv6 does not allow for fragmentation and reassembly at intermediate routers. It also does not have a checksum (which was deemed unnecessary once that Ethernet, UDP, TCP, and several other protocols already have checksums). Lastly, it does not have the variable length options field that IPv4 did.
The Internet is a set of autonomous systems which use a common protocol to communicate to each other what the current topology is. Within these autonomous systems, any routing algorithm might be used to define the best routes for traffic. In smaller, access networks, link-layer switches are more prevalent than routers. Lower-tier ISPs connect their autonomous systems to higher-tier ISPs in order to be able to reach all parts of the Internet.
An AS or autonomous system is a group of routers under the same administrative control. All the routers within an AS have to run the same routing algorithm and have some information about each other. One or more routers in an AS have the added task of forwarding packets to outside the AS. These routers are called gateway routers.
An ISP or Internet service provider is an organization that provides services for using the Internet. There are commercial, community-owned, non-profit, and privately owned ISPs.
Tier 1 ISPs almost never pay for IP transit, while tier 3 ISPs almost always pay for IP transit.
A point of presence (PoP) is where customer ISPs connect to provider ISPs.
Multi-homing occurs when a customer ISP connects itself to several providers in order to improve fault-tolerance or throughput.
In order to avoid paying for expensive provider ISP traffic, neighboring customer ISPs can exchange traffic directly. Just as tier 1 ISPs usually peer settlement-free, customer ISPs also usually peer settlement-free, that is, not paying for traffic.
An Internet exchange point (IXP) is a meeting point where multiple ISPs can peer together. They are usually set up by third-party companies.
The AS path can be made longer artificially by repeating parts of it. This way, through BGP, an AS can decide from which AS it would rather receive traffic. For instance, an AS A1 might tell one of its neighbors that going through A1 requires 10 more hops than it actually takes, therefore making other ASs more likely to use other neighbors to reach A1.
A BGP session that spans multiple ASs is called an external BGP (eBGP) session and a BGP session between routers in the same AS is called an internal BGP (iBGP) session. Certain attributes such as local preference are sent to iBGP peers but not to eBGP peers. Additionally, routes received from an eBGP peer can be advertised to both eBGP and iBGP peers. However, routes received from an iBGP peer cannot be advertised to other iBGP peers, only to eBGP peers. This is required to prevent loops.
TCP uses a byte stream abstraction, while UDP uses a datagram abstraction.
This is not specified. It is up to the implementation.
TCP keeps an adaptive calculation to determine the timeout. If the timeout is too short, it will cause unnecessary retransmissions. If the timeout is too long, the sender will be slow to react to losses.
There are no timers per segment, timers are per connection. The timer can be understood as the timer for the oldest segment which has not yet been acknowledged. When a timeout occurs, the segment is retransmitted and the timer is restarted.
The timeout is the estimated round-trip time plus a “safety margin”.
The estimated RTT () is an exponential weighted moving average. It uses a parameter and evaluates the following expression.
Typically, is 0.125.
In which is the measured time from segment transmission until acknowledgement receipt. This calculation ignores retransmissions.
The deviated RTT () is evaluated given by the following expression.
Typically, is 0.25. Larger values for will make deviations weight more into the timeout.
In both of these formulas, having larger values of the factor or makes the formula more sensitive to recent values. Conversely, smaller values make the metric to be more based on the past.
The timeout can be evaluated through the following formula.
Timeouts and three ACKs for the same segment. Retransmitting after three ACKs is known as fast retransmit.
In Fast Recovery, which is entered after three duplicate ACKs, the congestion window is halved and then grows linearly. After a timeout, we are back to the slow start, in which the congestion window has size = 1 MSS and grows exponentially.
The reasoning behind this is that multiple duplicate ACKs indicate a network capable of delivering some segments. However, a timeout indicates a more serious congestion or even failure of the network.
A variable controls the slow start threshold, which determines when the connection changes to AIMD. It is typically set to half the congestion window just before the loss event.
It refined Reno’s Fast Recovery. When selective acknowledgement is not being used and the algorithm enters Fast Recovery triggered by three duplicate acknowledgements, it does not know which segments to retransmit other than the first unacknowledged segment. NewReno gives special treatment to partial acknowledgements that might happen during Fast Recovery. It does not halve the congestion window if another segment was also not delivered, like Reno would, and will keep retransmitting the first unacknowledged segment until it leaves Fast Recovery. Because NewReno can send new packets at the end of the congestion window during Fast Recovery, high throughput is maintained during the Fast Recovery process. When it enters Fast Recovery, it stores the highest unacknowledged packet sequence number. When this sequence number is acknowledged, it returns to the congestion avoidance state.
Tries to deal with multiple packet losses as a single congestion event.
Splits site resources across different origins, in order to allow the client to open more than six TCP connections.
This caused less modular code and expensive cache invalidations when anything changed. It also slowed down execution, as the browser had to wait for the whole file to be downloaded.
Content prioritization introduced in HTTP/2 helps alleviate the need for this.
Preprocessing was required to make up the sprites. Also, the whole sprite bitmap had to be decoded at once, which was heavy on the client.
Prevents resources from being cached. Pays the overhead of Base64 encoding.
This is addressed by HTTP/2 through server push.
Binary framing was introduced in HTTP/2.
All binary frames have an 8-byte header with the fields length (16 bits), type (8 bits), flags (8 bits), 1 reserved bit, and the stream identifier (31 bits).
HTTP/2 performs header compression, in which both client and server maintain “header tables”. Therefore, only changes to the headers have to be sent through the network.
Can leverage large buffers and CDNs. The most resilient to latency.
More sensitive to latency than stored video.
The most sensitive to network oscillations, as the latency has to be kept low.
The client wants to have continuous playout of video. However, the network delay and its variation (jitter) make the time of arrival of frames variable. Client-side buffering alleviates this problem by storing more video than the barely minimum needed for continuous playout.
CDNs distribute content to cache servers located close to users, resulting in fast, reliable applications and Web services for the users.
More specifically, they maintain multiple points of presence with clusters of surrogate servers that store copies of identical content.
The CDN node may be selected as the one that is the fewest hops or the least number of network seconds away from the requesting client. It may also be the one with the highest availability in terms of server performance, so as to optimize delivery across local networks. It may even be optimized for cost, and the locations that are least expensive may be chosen instead. These two goals tend to align as the cheapest is usually the closest location.
Content outsourcing is the operation of moving data into a CDN.
Three common content outsourcing practices are the following.
Typically, Web content is grouped based on either correlation or access frequency and then replicated in units of content clusters. Content clustering can be based on user session statistics or on the site topology.