summaryrefslogtreecommitdiffstats
path: root/doc/trex_book.asciidoc
diff options
context:
space:
mode:
authorHanoh Haim <hhaim@cisco.com>2016-11-16 13:41:58 +0200
committerHanoh Haim <hhaim@cisco.com>2016-11-16 13:41:58 +0200
commite710eb541e8657f662c8bb2ef3252c42ad2544a6 (patch)
treec49bc5e9bf4668759481cd2953d360309638f3e0 /doc/trex_book.asciidoc
parent45f019c410bdc4a48f19458c655dc6d5232b43b9 (diff)
update ConnectX-4 performance numbers
Signed-off-by: Hanoh Haim <hhaim@cisco.com>
Diffstat (limited to 'doc/trex_book.asciidoc')
-rwxr-xr-xdoc/trex_book.asciidoc37
1 files changed, 35 insertions, 2 deletions
diff --git a/doc/trex_book.asciidoc b/doc/trex_book.asciidoc
index 41151eb7..fc69bdf4 100755
--- a/doc/trex_book.asciidoc
+++ b/doc/trex_book.asciidoc
@@ -2070,6 +2070,10 @@ The disadvantage is the OFED dependency.
The following distro were tested with TRex and OFED, others might work
* Ubuntu 14.04.3 LTS (GNU/Linux 3.19.0-25-generic x86_64)
+
+
+The following distro were tested and did *not* work for us
+
* Fedora 21 (3.17.4-301.fc21.x86_64)
==== Install OFED
@@ -2399,14 +2403,40 @@ NIC with two ports will work better from performance prospective, so it is bette
* link:https://trex-tgn.cisco.com/youtrack/issue/trex-260[64B performance issue]
* link:https://trex-tgn.cisco.com/youtrack/issue/trex-261[Latency issue]
* link:https://trex-tgn.cisco.com/youtrack/issue/trex-262[Statful RX out of order]
+* link:https://trex-tgn.cisco.com/youtrack/issue/trex-273[Fedora 21 & OFED 3.4.1]
+
+
+==== Performance Cycles/Packet ConnectX-4 vs Intel XL710
+
+For version TRex v2.11, this is the comparison results between XL710 and ConnectX-4 for various scenarios
+
+.Stateless MPPS/Core [Perlimeniary]
+image:images/xl710_vs_mlx5_64b.png[title="Stateless 64B"]
+
+.Stateless Gb/Core [Perlimeniary]
+image:images/xl710_vs_mlx5_var_size.png[title="Stateless variable size packet"]
+
+*Comments*::
+
+1. For Stateless 64B profiles ConnectX-4 cost 50-90% more cycles per packet (it is actually even more because there is the TRex scheduler overhead) - it means that in the worst case scenario you will need x2 CPU for the same total MPPS
+2. For Stateless/Stateful 256B profiles, ConnectX-4 cost half of the cycles per packets. ConnectX-4 probably can handle in a better way chained mbuf (scatter gather).
+3. In Average Stateful senario ConnectX-4 will be slightly better.
+4. MLX5 can reach ~90MPPS while XL710 limited to 35MPPS
+
+
+[NOTE]
+=====================================
+There is a task to automate the production of this reports
+=====================================
=== Cisco VIC support
anchor:ciscovic_support[]
+* Supported from TRex version v2.11
* Only 1300 series Cisco adapter
-* Firmware version 2.0(13) for UCS C-series servers
+* Firmware version 2.0(13) for UCS C-series servers, GA in Febuary 2017.
* Firmware version 3.1(2) for blade servers supports more filtering capabilities.
* The feature can be enabled via Cisco CIMC or USCM with the 'advanced filters' radio button. When enabled, the these additional flow director modes are available:
RTE_ETH_FLOW_NONFRAG_IPV4_OTHER
@@ -2420,5 +2450,8 @@ anchor:ciscovic_support[]
==== Limitation/Issues
* Stateless per stream statistic is not supported yet
-* link:https://trex-tgn.cisco.com/youtrack/issue/trex-265[flow-director]
+* link:https://trex-tgn.cisco.com/youtrack/issue/trex-265[flow-director]
+* link:https://trex-tgn.cisco.com/youtrack/issue/trex-272[QSFP+ issue]
+
+