94
Other(19.9%)
Unaccounted(5.2%)
Sync(5.2%)
NFS(6.1%)
Network(11.9%)
Bcopy/Bcmp(10.1%)
MemMgt(4.7%)
Waiting(10.8%)
UFS(10.7%)
SCSI-driver(15.4%)
Kernel Time (SCSI)
% time in sub-system
Other(26.1%)
Unaccounted(4.0%)
Sync(6.5%)
NFS(3.3%)
Network(17.1%)
Bcopy/Bcmp(13%)
MemMgt(5.7%)
Waiting(4%)
UFS(14%)
RAID-driver(5.1%)
Kernel Time (RAID)
% time in sub-system
(a)
(b)
Figure 5.9:
Time Breakdown Near Peak Op/sec
These charts show the percentage of time spent in each sub-system when operating near the satura-
tion point. Measurements for the graph on the left were taken on the SCSI system at 1000 ops/sec,
while measurements for the graph on the right were taken on the RAID system at 1400 ops/sec. The
area of each chart shows the relative time per operation, including idle time (i.e., waiting on I/O),
of 1 msec for the SCSI and 714
à
s for the RAID.
5.5.4
Bulk Gap
We choose to examine sensitivity to bulk Gap,
á
, as opposed to the per-message rate
â
.
First, networking vendors often tout per-byte bandwidth as the most important metric in comparing
networks. Using our apparatus we can quantify its sensitivity (and thus importance). Secondly, for
the SPECsfs benchmark,
â
is quite low (in the 1000’s msg/sec range) and is easily handled by most
networks.
Unlike overhead, which is incurred on every message, sensitivity to Gap is incurred only
if the data rate exceeds the Gap. Only if the processor sends data in an interval smaller than that
specified by
á
will it stall. The clients and server could potentially ignore
á
entirely. Recall the
burst vs. uniform models for gap presented in Section 3.2.2. At one extreme, if all data is sent at a
uniform rate that is less than
á
we will not observe any sensitivity to Gap. At the other extreme, if
all data is sent in bursts then we would observe maximum sensitivity to Gap.
Because SPECsfs sends messages at a controlled rate, we would expect that message in-
tervals are not bursty and the benchmark should be quite insensitive to Gap. Figure 5.10 shows that
this is indeed the case. Only when the bandwidth (
ã
ä
) falls from a baseline of 26 MBs to a mere 2.5
MB/s do we observe any sensitivity to
á
. We are thus assured that the SFS benchmark is not bursty.
Measured production environments, however, are quite bursty [49, 66]. Our measured sen-