--- discussion.tex      (revision 2510)
+++ discussion.tex      (working copy)
@@ -404,16 +404,14 @@
 Because the developed system is intended for large scale installations
 with multiple machines, we have designed the protocols and system
 architecture so that it should be able to scale well to a large number
-of hosts. In our tests we see a linear increase in completion time
-when scaling up to 50 hosts.
+of hosts. We have tested the simultaneous configuration of up to 50 hosts.
 
-If the system truly scales unlimitedly, we would expect to see a
-constant completion time with any number of hosts. Since we see a
-linear increase, this must be explained by increased load on some
+As the number of hosts increase, so does the total configuration time.
+This must be explained by increased load on some
 combination of the involved services. We can rule out the database
-server as a bottleneck, since only 1 query is made for each configured
-node. Even if the server could only handle 50 queries per second, it
-would still account for less than one percent of the time taken.
+server as a bottleneck. Only 1 query is made for each configured node,
+and we have measured such a query to take no more than 40ms, including
+connecting to the database.
 
 The MetaConfig server can be ruled out for the same reason. In
 bootstrap mode it performs only a few simple operations, which
@@ -424,7 +422,7 @@
 Even with the bottleneck present in the system, it still displays
 reasonable scalability. We have seen that as we go from 1 host to 50
 hosts, average configuration time only increases by about 75\%. As we
-can see from the graph we have a smooth linear progression from 10
+can see from the graph we have a smooth progression from 10
 hosts up to 50 hosts, with no signs of polynomial or exponential
 growth. The spikes in the graph at below 10 hosts are likely caused by
 cache locality issues or measurement uncertainty.
@@ -432,7 +430,7 @@
 In our test, we have configured up to 50 hosts simultaneously. In a
 real-world scenario, many more than 50 hosts would have to be present
 in a data center for that many of them to be configured at the same
-time. For instance, if we consider 10 minutes to be the longest
+time. Also, if we consider 10 minutes to be the longest
 average configuration time we are willing to accept, we can predict
 that our system should be able to handle approximately 120 hosts,
 based on the performance levels at 50 hosts. Having to buy an extra