Changes between Version 2 and Version 3 of ConcurrentHashTable

Timestamp:

2012-08-08T11:23:23Z (13 years ago)

Author:

Adam Hraska

Comment:

rcu read side scalability benchmark description

ConcurrentHashTable

@@ -95 +95 @@
 visible in the figures without zooming ;-).
 
+=== RCU read side scalability ===
+[[Image(r1589-list-read.png)]]
 
+First, we examine the overhead of rcu read sections compared to acquiring a spinlock. 
+The figure above shows the number of traversals of a five element immutable list
+depending on the number of threads/cpus used. More is better ;-).
+- //ideal// - the list was accessed without any synchronization whatsoever
+- //a-rcu// - each list traversal was protected by A-RCU 
+- //podzimek-rcu// - protected by the preemptible modification of Podzimek's RCU
+- //spinlock// - guarded by an ordinary preemption disabling spinlock
+
+A-RCU fares the best and has optimal scaling. On the other side, spinlock presents 
+negative scaling (ie the more cpus you throw at it, the slower it is). Podzimek's RCU 
+scales perfectly but has a greater base cost compared to A-RCU. In particular, 
+Podzimek-RCU's base cost is on par with a spinlock when running on a single cpu while 
+A-RCU's base cost is significantly lower than both Podzimek-RCU's and spinlock base cost.
+
+To reproduce these results, switch to the kernel console and run:
+{{{
+chtbench 2 1 0 -w
+chtbench 2 2 0 -w
+chtbench 2 3 0 -w
+chtbench 2 4 0 -w
+chtbench 3 1 0 -w
+chtbench 3 2 0 -w
+chtbench 3 3 0 -w
+chtbench 3 4 0 -w
+chtbench 4 1 0 -w
+chtbench 4 2 0 -w
+chtbench 4 3 0 -w
+chtbench 4 4 0 -w
+}}}
+
+[[Image(r1589-list-upd.png)]]
+[[Image(r1589-list-upd-trim.png)]]
+[[Image(r1589-ht-lookup.png)]]
+[[Image(r1589-ht-upd.png)]]