1 files changed, 69 insertions, 0 deletions
diff --git a/docs/aboutvpp/scalar-vs-vector-packet-processing.rst b/docs/aboutvpp/scalar-vs-vector-packet-processing.rst
new file mode 100644
index 00000000000..173efecebc5
--- /dev/null
+++ b/docs/aboutvpp/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.