Differences

This shows you the differences between two versions of the page.

--- documentation:technical_docs:performance [2019/06/27 01:37] – [Where is the bottleneck ?] olivier
+++ documentation:technical_docs:performance [2020/01/18 01:02] – [Polling mode] olivier
@@ Line 27: / Line 27: @@
   * [[http://www.telematica.polito.it/oldsite/courmayeur06/papers/06-A.2.1.pdf|RFC2544 Performance Evaluation for a Linux Based Open Router]] (2006, June)
   * [[http://data.guug.de/slides/lk2008/10G_preso_lk2008.pdf|Towards 10Gb/s open-source routing]] (2008): Include an hardware comparison between a "real" router and a PC.
+  * [[https://wiki.fd.io/images/7/7b/Performance_Consideration_for_packet_processing_on_Intel_Architecture.pptx|Performance consideration for packet processing on Intel Architecture (ppt)]]
 ==== FreeBSD ====
@@ Line 90: / Line 91: @@
 Avoid NUMA architecture but prefer a CPU in only one package with maximum core (8 or 16).
-If you are using NUMA, check that inbound/outbound NIC queues are correctly mapping to the same package.
+If you are using NUMA, you need to check that inbound/outbound NIC queues are correctly bind to their local package to avoid useless QPI crossing.
 === Network Interface Card ===
@@ Line 104: / Line 105: @@
 ==== Choosing good FreeBSD release ====
-Before tuning, you need to use the good FreeBSD version.
+Before tuning, you need to use the good FreeBSD version... this mean a recent FreeBSD -head.
-This mean a FreeBSD -head version older than r309257 (Andrey V. Elsukov 's improvement: Rework ip_tryforward() to use FIB4 KPI) backported to FreeBSD 11-stable r310771 (MFC to stable).
-BSDRP since version 1.70 is using a FreeBSD 11-stable (r312663) that includes this improvement.
+BSDRP is currently following FreeBSD 12-stable branch, to try to have a mix between recent features and stability.
+==== Disabling Hyper Threading (on specific CPU only) ====
-{{documentation:technical_docs:2016-performance-evolution.png|2016 Forwarding performance evolution of FreeBSD -head on a 8 core Atom}}
+By default a multi-queue NIC drivers create one queue per core.
+But on some older CPU (like Xeon E5-2650 V1) those logical cores didn't help at all for managing interrupts generated by high speed NIC.
-For better (and linear scale) performance there is the [[https://svnweb.freebsd.org/base/projects/routing/|projects/routing]] too [[http://blog.cochard.me/2015/09/receipt-for-building-10mpps-freebsd.html|that still give better performance]].
-==== Disabling Hyper Threading ====
-Disable Hyper Threading (HT): By default, lot's of multi-queue NIC drivers create one queue per core.
-But "logical" cores didn't help at all for managing interrupts generated by high speed NIC.
 HT can be disabled with this command:
@@ Line 123: / Line 118: @@
 </code>
-Here is an example on a 8cores x hardware threads Intel CPU and 10G Chelsio NIC:
+Here is an example on a Xeon E5 2650 (8c,16t) and 10G Chelsio NIC where it improve performance by disabling HT:
 <code>
-x HT-enabled-8rxq(default).packets-per-seconds
+x HT-enabled-8rxq(default): inet packets-per-second forwarded
-+ HT-enabled-16rxq.packets-per-seconds
++ HT-enabled-16rxq: inet packets-per-second forwarded
-* HT-disabled.packets-per-seconds
+* HT-disabled-8rxq: inet packets-per-seconds forwarded
 +--------------------------------------------------------------------------+
 |                                                                        **|
@@ Line 150: / Line 145: @@
 </code>
-There is a benefit of about 24% to disable hyper threading.
+There is a benefit of about 24% to disable hyper threading on this old CPU.
+But here is another example where there is a benefit to kept it enabled (and with the NIC configured to uses all the treads) on Xeon E5 2650L (10c, 20t):
+<code>
+x HT on, 8q (default): inet4 packets-per-second forwarded
++ HT off, 8q: inet4 packets-per-second forwarded
+* HT on, 16q: inet4 packets-per-second forwarded
++--------------------------------------------------------------------------+
+|x x              ++                                                   *  *|
+|x xx            +++                                                 * *  *|
+||AM|            |A_|                                                |_MA_||
++--------------------------------------------------------------------------+
+    N           Min           Max        Median           Avg        Stddev
+x   5       4265579     4433699.5     4409249.5     4359580.3       81559.4
++   5       5257621       5443012       5372493     5372693.5     73316.243
+Difference at 95.0% confidence
+.01311e+06 +/- 113098
+.2388% +/- 2.94299%
+        (Student's t, pooled s = 77547.4)
+*   5       8566972       8917315     8734750.5     8769616.1     147186.74
+Difference at 95.0% confidence
+.41004e+06 +/- 173536
+.157% +/- 5.21388%
+        (Student's t, pooled s = 118987)
+</code>
 ==== fastforwarding ====
@@ Line 235: / Line 255: @@
 {{documentation:technical_docs:entropy_source_impact.png|Impact of disabling some entropy source on FreeBSD forwarding performance}}
-==== Polling mode ====
+==== Polling mode (very old hardware) ====
 Polling can be used in 2 cases:
@@ Line 746: / Line 766: @@
 </code>
-On this case the bootleneck is just the network stack.
+On this case the bootleneck is just the network stack (most of the time spend into function ip_findroute called by ip_tryforward).
 == CPU cycles spent ==
@@ Line 764: / Line 784: @@
 less /data/pmc.stacks
 </code>
+=== Lock contention source ===
+To identifying lock contention source (like if function lock_delay or __mtx_lock_sleep was quite high from the pcm output), you can try to search which lock is contended and why with lockstat.
+You can generate 2 output:
+  * contented locks broken down by type: <code>lockstat -x aggsize=4m sleep 10 > lock-type.txt</code>
+  * stacks associated with the lock contention to identify the source: <code>lockstat -x aggsize=4m -s 10 sleep 10 > lock-stacks.txt </code>