LDD Today


[back to "Optimizing server performance: CPU scalability"]
CPU scalability test results (sidebar)
The following charts show the results of our tests. The first chart shows CPU scalability using RAID0 and the second chart shows CPU scalability using RAID5. For each test we used the R5.0 maildb as our workload. We ran each test with a user load of 1800 users.

NotesMarks are the Domino throughput metric of transactions-per-minute (tpm). They measure the amount of work the Domino server performs. The values are only comparable when the same workload is being applied to a server (as in our test scenario). The NotesMark values indicate that the server did the same amount of work regardless of the number of CPUs (the minor differences are insignificant).

The Percentage of the CPU Used column shows the percentage of the total CPU utilization, which is the sum of the percentage of user utilization, plus the percentage of system CPU utilization. These percentages come from vmstat, a standard UNIX system utility that provides virtual memory usage statistics. The Scan Rate column shows the pages scanned per second by the virtual memory system. This corresponds to the virtual memory activity as reported by the Scan Rate column in vmstat. The Disk Service Time column shows the average disk service time in milliseconds from the sar disk activity report. This report contains the average disk I/O operation time. Sar, the system activity reported, is a UNIX utility which samples cumulative activity counters in the operating system at the interval specified. This shows you the average time for servicing a read/write disk request.

Table 1: CPU scalability using RAID0
Number of CPUs
NotesMark
(Domino Transactions)
Response
time (ms)
Percentage of the CPU used
Scan Rate
(Virtual Memory)
Disk Service Time (ms)
4
3331
5027
51
75
14
8
3638
100
25
75
14
12
3628
95
14
75
14

Table 2: CPU scalability using RAID5
Number of CPUs
NotesMark
(Domino Transactions)
Response
time (ms)
Percentage of the CPU used
Scan Rate
(Virtual Memory)
Disk Service Time (ms)
4
3149
6730
60
100
35
8
3596
250
28
100
35
12
3598
246
17.5
100
35
When we compared the information in the two tables, we observed that RAID5 had a 30 percent, to over 200 percent, relative impact on response time. You can see this in the following graph. When we test response time, we often refer to both the absolute and the relative response time of a Domino server. The absolute response time is the measured time reported by Notesnum.exe. Notesnum is a utility that aggregates the response time from each Notes client simulator (we used up to five) to produce an overall response time average for all the simulated users. The relative response time is the difference in time between the series data point (in this case, the relative difference between the response time at four CPUs using RAID0 and four CPUs using RAID5).

Relative RAID Rsp Time Impact

However, as you can see in the following graph, the absolute response times were much better at the higher load values. The small-value response times (for example, 100ms versus 250ms) produced large relative response time differences (so it's important to assess both the absolute and relative values when making decisions). You should interpret this data as indicating that RAID5 did have an impact, but not a substantial impact. The increased data reliability provided by RAID5 is worth the increase in resource utilization. End users would not perceive any response time difference between 100ms and 250ms.

RAID Rsp Time Impact