A LowPower Network Search Engine Based on Statistical Partitioning

1
A Low-Power Network Search Engine Based on
Statistical Partitioning

• Taskin Kocak and Faysal Basci
• Dept. of Electrical and Computer Engineering
• University of Central Florida, Orlando, FL 32816
• tkocak_at_cpe.ucf.edu

2
Introduction

• Stringent memory access and search speed
  requirements are two of the main bottlenecks in
  wire speed packet processing.
• Most viable search engines are implemented in
  Content Addressable Memories (CAM).
• In networking applications, CAMs are generally
  used in VLAN
• lookup tables (layer-2), packet forwarding lookup
  tables (layer-3), session lookup table (layer-4)
  and filtering (layers 5-7)
• CAMs have high operational speed advantage over
  other memory search algorithms, such as
  look-aside tag buffers, binary or tree based
  searches.
• However, this performance advantage comes with a
  price of higher silicon area, and higher power
  consumption (3-5 W/chip).

3
CAM Architecture
4
IP lookups

• IP lookups are based on classless inter-domain
  routing (CIDR) scheme.
• The IP address space is broken into line
  segments.
• Each line segment is described by a prefix.
• A prefix is of the form x/y where x indicates
  the prefix of all addresses in the line segment,
  and y indicates the length of the segment.
• e.g. The prefix 134.26/16 represents the line
  segment containing addresses in the range
  134.26.0.0 134.26.255.255.
• The most specific route longest prefix match

5
TCAMs in IP forwarding

• TCAMs dont care storage capability is
  favorable in lookup engines.
• e.g., for a 24-bit prefix, the last 8 bits will
  be x
• Routing table entries stored in TCAMs are
  ordered according to their prefixes
• In the case of multiple matches priority encoder
  will choose the one with the lowest address,
  which is the longest matching prefix

6
32-bit IP Prefix Distribution
7
Partitioned TCAM
Perform a search in TCAM1 If there is a match
DONE otherwise search TCAM2 If
there is a match DONE otherwise
issue mismatch end
8
Power Consumption
PTCAM PSTATIC PCLOCK PMATCH PMISS
PX Here, we are concerned about the dynamic power
consumption caused by the search (comparison)
operation.
where ?0 represent the probability the search
word has a match in the whole table, and ?1
represent the fraction of entries that resides in
TCAM1
9
Power Consumption (cont.)
Assuming that the probability of match for a
single cell is 0.5
10
Experimental Results dynamic power
Example TCAM circuit is implemented in Cadence
for TSMC 0.18-µm CMOS. Simulations are run at 100
MHz. PMATCH 397 nW PMISS 515 nW Px
336 nW
11
Experimental Results dynamic power
12
Experimental Results powerlatency
EL2-?0 L search latency in clock cycles
13
Power consumption with different miss rates for
Telstra
14
Experimental Results
Expected power savings and average latency when
partitioning into three TCAMs
15
Reconfigurable Architecture
16
Conclusions

• We presented a TCAM partitioning scheme, which
  utilizes statistical distribution of prefixes.
• We showed that indeed the partitioning helps
  reducing the power consumption in IP lookup
  applications.
• Partitioning into two and three, reduce power
  consumption considerably, whereas further
  partitioning shows little improvement

17
ANCHOR workshop at ISCA 2004
18
Dynamic power