IDT PCI Express 89HPES32NT24xG2, User Manual, Page: 89 / 744

IDT Switch Core
PES32NT24xG2 User Manual 4 - 11 January 30, 2013
Notes
Request Metering
Request metering may be used to reduce congestion in PCI Express switches caused by a static rate
mismatch. Request metering is available on all PES32NT24xG2 switch ports but is disabled by default. The
DMA function also has a mechanism to meter requests that it generates. This mechanism operates inde-
pendently from the mechanism described in this section. Refer to section DMA Request Rate Control on
page 15-22 for details.
A static rate mismatch is a mismatch in the capacity of the path from a component injecting traffic into
the fabric (e.g., a Root Complex) and the ultimate destination (e.g., an Endpoint). An example of a static
rate mismatch in a PCI Express fabric is a x8 root injecting traffic destined to a x1 endpoint. PCI Express
fabrics are typically no more than one switch deep. Therefore, static rate mismatches typically occur within
a switch due to asymmetric link rates.
Figure 4.3 illustrates the effect of congestion on PCI Express fabric caused by a static rate mismatch. In
this example, there are two endpoints issuing memory read requests to a root. Endpoint A has a x1 link to
the switch, while endpoint B and the root complex have a x8 link.
Memory read request TLPs are three or four DWords in size. A single memory read request may result
in up to 4 KB of completion data being returned to the requester. Depending on system architecture and
configured maximum payload size, this completion data may be returned as a single completion TLP or
may be returned as a series of small (e.g., 64B data) TLPs.
Consider an example where endpoints A and B are injecting read request to the root at a high rate and
the root is able to inject completion data into the fabric at a rate higher than which may be supported by
endpoint A’s egress link. The result is that the endpoint A’s EFB and the root’s IFB may become filled with
queued completion data blocking completion data to endpoint B.
5.0 x8 2.5 x8, x4, x2, x1 Always
5.0 x8, x4, x2, x1 Always
x4 2.5 x8, x4, x2, x1 Always
5.0 x8 At least 50% of packet is in IFB
x4, x2, x1 Always
x2 2.5 x8 At least 50% of packet is in IFB
x4, x2, x1 Always
5.0 x8 At least 75% of packet is in IFB
x4 At least 50% of packet is in IFB
x2, x1 Always
x1 2.5 x8 At least 75% of packet is in IFB
x4 At least 50% of packet is in IFB
x2, x1 Always
5.0 x8 Never (100% of packet is in IFB)
x4 At least 75% of packet is in IFB
x2 At least 50% of packet is in IFB
x1 Always
Ingress
Link
Speed
(GT/s)
Ingress
Link
Width
Egress
Link
Speed
(GT/s)
Egress
Link
Width
Conditions for
Cut-Through
Table 4.5 Conditions for Cut-Through Transfers (Part 2 of 2)