summaryrefslogtreecommitdiffstats
path: root/docs/whatisvpp/scalar-vs-vector-packet-processing.rst
diff options
context:
space:
mode:
authorJohn DeNisco <jdenisco@cisco.com>2019-11-06 10:58:28 -0800
committerDave Barach <dave@barachs.net>2019-11-06 16:15:49 -0500
commitc96d618a5dd96e3a40d59860d2cdb9d5c6b71d11 (patch)
tree74a19b9b8364bf56dceced8ec982c6fbb7ddb8e4 /docs/whatisvpp/scalar-vs-vector-packet-processing.rst
parent340c15c6ed34ce60c821b5260fec3eb11d65dcb7 (diff)
docs: Rewrite the what is VPP (first) section, also fix the build
Signed-off-by: John DeNisco <jdenisco@cisco.com> Change-Id: Ifb558171f8976a721703e74afea997d006273b5f Signed-off-by: Dave Barach <dave@barachs.net>
Diffstat (limited to 'docs/whatisvpp/scalar-vs-vector-packet-processing.rst')
-rw-r--r--docs/whatisvpp/scalar-vs-vector-packet-processing.rst69
1 files changed, 69 insertions, 0 deletions
diff --git a/docs/whatisvpp/scalar-vs-vector-packet-processing.rst b/docs/whatisvpp/scalar-vs-vector-packet-processing.rst
new file mode 100644
index 00000000000..ffa54a3f306
--- /dev/null
+++ b/docs/whatisvpp/scalar-vs-vector-packet-processing.rst
@@ -0,0 +1,69 @@
+.. _scalar_vector:
+
+==================================
+Scalar vs Vector packet processing
+==================================
+
+FD.io VPP is developed using vector packet processing, as opposed to
+scalar packet processing.
+
+Vector packet processing is a common approach among high performance packet
+processing applications such FD.io VPP and `DPDK <https://en.wikipedia.org/wiki/Data_Plane_Development_Kit>`_.
+The scalar based approach tends to be favoured by network stacks that
+don't necessarily have strict performance requirements.
+
+**Scalar Packet Processing**
+
+A scalar packet processing network stack typically processes one packet at a
+time: an interrupt handling function takes a single packet from a Network
+Interface, and processes it through a set of functions: fooA calls fooB calls
+fooC and so on.
+
+.. code-block:: none
+
+ +---> fooA(packet1) +---> fooB(packet1) +---> fooC(packet1)
+ +---> fooA(packet2) +---> fooB(packet2) +---> fooC(packet2)
+ ...
+ +---> fooA(packet3) +---> fooB(packet3) +---> fooC(packet3)
+
+
+Scalar packet processing is simple, but inefficient in these ways:
+
+* When the code path length exceeds the size of the Microprocessor's instruction
+ cache (I-cache), `thrashing
+ <https://en.wikipedia.org/wiki/Thrashing_(computer_science)>`_ occurs as the
+ Microprocessor is continually loading new instructions. In this model, each
+ packet incurs an identical set of I-cache misses.
+* The associated deep call stack will also add load-store-unit pressure as
+ stack-locals fall out of the Microprocessor's Layer 1 Data Cache (D-cache).
+
+**Vector Packet Processing**
+
+In contrast, a vector packet processing network stack processes multiple packets
+at a time, called 'vectors of packets' or simply a 'vector'. An interrupt
+handling function takes the vector of packets from a Network Interface, and
+processes the vector through a set of functions: fooA calls fooB calls fooC and
+so on.
+
+.. code-block:: none
+
+ +---> fooA([packet1, +---> fooB([packet1, +---> fooC([packet1, +--->
+ packet2, packet2, packet2,
+ ... ... ...
+ packet256]) packet256]) packet256])
+
+This approach fixes:
+
+* The I-cache thrashing problem described above, by amortizing the cost of
+ I-cache loads across multiple packets.
+
+* The inefficiencies associated with the deep call stack by receiving vectors
+ of up to 256 packets at a time from the Network Interface, and processes them
+ using a directed graph of node. The graph scheduler invokes one node dispatch
+ function at a time, restricting stack depth to a few stack frames.
+
+The further optimizations that this approaches enables are pipelining and
+prefetching to minimize read latency on table data and parallelize packet loads
+needed to process packets.
+
+Press next for more on Packet Processing Graphs.