Free Trial

Safari Books Online is a digital library providing on-demand subscription access to thousands of learning resources.

Share this Page URL
Help

15.1 INTERNAL FLOW CONTROL > 15.1.2 Rescuing Reliability - Pg. 365

15.1 Internal Flow Control 365 But allocating the full pipe size to all classes at the same time is obvious waste (P1) because if every class were to send cells at the same time, each by itself would get only a fraction of the link throughput. Thus it makes sense to share buffers. The simplest approach to buffer sharing is to divide the buffer space physically into a common pool together with a private pool for each class. A naive method to do so would mark data cells and credits as belonging to either the common or the private pools to prevent interference between classes. The naive scheme also requires additional complexity to guarantee that a class does not exceed, say, a pipe size worth of buffers. An elegant way to achieve the allow buffer sharing without marking cells is described in Ozveren et al. [OSV94]. Conceptually, the entire buffer space at the receiver is partitioned so that each class has a private pool of Min buffers; in addition there is a common pool of size (B - N Min) buffers, where N is the number of classes and B is the total buffer space. Let Max denote the pipe size. The protocol runs in two modes: congested and uncongested. When congested, each class is restricted to Min outstanding cells; when uncongested, each class is allowed the presumably larger amount of Max outstanding cells. All cell buffers at the downstream node are anonymous; any buffer can be assigned to the incoming cells of any class. However, by carefully restricting transitions between the two modes, we can allow buffer sharing while preventing deadlock and cell loss. To enforce the separation between private pools without marking cells, the sender keeps track of the total number of outstanding cells S, which is the number of cells sent minus the number of credits received. Each class i also keeps track of a corresponding counter S i , which is the number of cells outstanding for class i. When S < N · Min (i.e., the private pools are in no danger of depletion), then the protocol is said to be uncongested and every class i can send as long as S i Max. However, when S N · Min, the link is said to be congested and each class is restricted