Released
PMC-Sierra, Inc.
PM9311/2/3/5 ETT1™ CHIP SET
Data Sheet
PMC-2000164
ISSUE 3
ENHANCED TT1™ SWITCH FABRIC
cost savings can be achieved by using fewer Crossbar devices. A Crossbar can be configured to be in an
8- 16- or 32-port core, as shown in Table 1.
Table 1. Crossbar Configurations
Number of Crossbars required Number of Crossbars required
Number of OC-192c ports
(non redundant)
(redundant)
1 to 8
9 to 16
17 to 32
3 + 1 Flow Control
6 + 1 Flow Control
12 + 2 Flow Control
6 + 2 Flow Control
12 + 2 Flow Control
24 + 4 Flow Control
The reduction in the number of crossbars is achieved by having a single port use more than one
connection on each Crossbar. In a 32 port system, each port will use one connection to each Crossbar
device; in an 8 port system, each port will use four connections on each Crossbar. (This DOES NOT apply
to Flow Control Crossbars; each port will always use only one connection.)
1.3 PRIORITIZED BEST-EFFORT QUEUE MODEL
The ETT1 switch core has been designed to maximize cell throughput while keeping latency small and
providing “fair” throughput among all ports. The ETT1 core supports four levels of strict priority. Level 0 is
the highest priority and level 3 the lowest. All conditions being equal, the core will always forward a higher
priority cell before a lower priority cell. Furthermore, the core will be fair to all inputs. This fairness is very
difficult to quantify and is only meaningful over time periods of many cells. Over short time periods (tens or
perhaps hundreds of cells) there may be temporary unfairness between ports.
A cell of a lower priority may pass a cell of a higher priority under certain conditions:
1. Since the scheduling pipeline is shorter for lower priorities, a cell of a lower priority may emerge
from the switch sooner than a cell of a higher priority if the lower-priority cell is sent immediately
after the higher-priority cell.
2. Hole requests for a higher priority can allow cells of a lower priority to pass cells of the higher
priority.
3. If the linecard responds to LCS grants faster than required to meet the system round trip time
constraint, then multicast cells of a lower priority may pass unicast cells of a higher priority, due to
a (programmable) delay in the EPP’s “Internal Delay Matching Adjustments” register.
4. In subport mode only: multicast output queues (VIQs) are shared by subports, so when a subport
is oversubscribed for an extended period of time, blocking may result on all subports. This
blocking of multicast can allow unicast cells of the same priority to pass.
In this section we describe the queueing model provided by the ETT1 switch core for this prioritized
best-effort traffic. We start by considering only OC-192c ports. The following sections describe how the
model differs for OC-48c ports.
30
PROPRIETARY AND CONFIDENTIAL TO PMC-SIERRA, INC., AND FOR ITS CUSTOMERS’ INTERNAL USE