Contents Telektronikk - Telenor
Contents Telektronikk - Telenor
Contents Telektronikk - Telenor
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
Time [ms]<br />
Time [ms]<br />
3.6<br />
3.2<br />
2.8<br />
2.4<br />
2.0<br />
1.6<br />
1.2<br />
.8<br />
.4<br />
0<br />
3.6<br />
3.2<br />
2.8<br />
2.4<br />
2.0<br />
1.6<br />
1.2<br />
.8<br />
.4<br />
0<br />
0<br />
0<br />
SBA-200/2.2.6 - driver<br />
SBA-200/2.2.6 - total<br />
SBA-100/2.2.6 - driver<br />
SBA-100/2.2.6 - total<br />
SBA-100/2.0 - driver<br />
SBA-100/2.0 - total<br />
SBA-200/2.2.6 - driver<br />
SBA-200/2.2.6 - total<br />
SBA-100/2.2.6 - driver<br />
SBA-100/2.2.6 - total<br />
not fully representative for performance<br />
when the protocol operates like stop-andgo.<br />
In order to explain the Sparc2 and<br />
Sparc10 relative performance for smaller<br />
window sizes when using SBA-100<br />
adapters, we measured the throughput<br />
with TCP operating as a stop-and-go protocol<br />
for these configurations. This was<br />
achieved by setting the window size<br />
equal to the user data size. The throughput<br />
under such a behavior is shown in<br />
Figure 5 for both host architectures. Up<br />
to a user data size between 2048 and<br />
2560 bytes the slower Sparc2 has a higher<br />
throughput. This is most likely due to<br />
the longer Sbus access time combined<br />
with the relatively longer interrupt time.<br />
4 Initial throughput<br />
measurements<br />
2048 4096 6144 8192<br />
Segment size [byte]<br />
(a) Sparc2 send times<br />
2048 4096 6144 8192<br />
Segment size [byte]<br />
(c) Sparc10 send times<br />
This section presents the initial throughput<br />
measurements for the reference architecture,<br />
the Sparc2 with the SBA-100<br />
network adapter with the first network<br />
driver version, i.e. version 2.0. Figure 6<br />
(a) presents the measured TCP end-toend<br />
memory-to-memory throughput for<br />
Time [ms]<br />
Time [ms]<br />
3.6<br />
3.2<br />
2.8<br />
2.4<br />
2.0<br />
1.6<br />
1.2<br />
.8<br />
.4<br />
0<br />
3.6<br />
3.2<br />
2.8<br />
2.4<br />
2.0<br />
1.6<br />
1.2<br />
.8<br />
.4<br />
0<br />
0<br />
0<br />
SBA-200/2.2.6 - driver<br />
SBA-200/2.2.6 - total<br />
SBA-100/2.2.6 - driver<br />
SBA-100/2.2.6 - total<br />
SBA-100/2.0 - driver<br />
SBA-100/2.0 - total<br />
SBA-200/2.2.6 - driver<br />
SBA-200/2.2.6 - total<br />
SBA-100/2.2.6 - driver<br />
SBA-100/2.2.6 - total<br />
Figure 4 Segment send and receive processing times<br />
different user data sizes and different<br />
window sizes. The throughput dependent<br />
on window size for a fixed user data size<br />
of 8192 bytes is presented in Figure 6<br />
(b), and the corresponding CPU utilization<br />
in Figure 6 (c). Both the user data<br />
size and the window size affect measured<br />
throughput. From the graphs it is clear<br />
that increasing the TCP window size<br />
above 32 kbytes has no increasing effect<br />
on the throughput. On the contrary, such<br />
an increase in window size results in a<br />
significantly varying throughput dependent<br />
on user data size. The reason is cell<br />
loss at the receiving host which causes<br />
TCP packet loss and retransmissions of<br />
lost packets.<br />
In general, the receive operation is<br />
known to be more time consuming than<br />
the transmit operation. At the receive<br />
side, there are among other things<br />
demultiplexing and scheduling points<br />
and interrupts which do not occur at the<br />
send side. In addition, reading from<br />
memory is more time consuming than<br />
writing to memory. Clearly, this is evident<br />
from the driver processing times<br />
presented in Figure 4. Furthermore, the<br />
processor (CPU) utilization histograms in<br />
2048 4096 6144 8192<br />
Segment size [byte]<br />
(b) Sparc2 receive times<br />
2048 4096 6144 8192<br />
Segment size [byte]<br />
(d) Sparc10 receive times<br />
Figure 6 (c) show the receiver to be<br />
heavier loaded than the sender. When the<br />
TCP connection starts losing packets due<br />
to cell loss at the receiver, both the<br />
throughput and CPU utilization are<br />
degraded. TCP employs positive<br />
acknowledgment with go-back-n retransmission.<br />
The sender fills the window,<br />
and if every byte is not acknowledged, it<br />
relies on retransmission timers to go off<br />
before it starts sending at the point of the<br />
Throughput |Mbit/s]<br />
30<br />
20<br />
10<br />
Sparc2<br />
Sparc10<br />
0<br />
0 1024 2048 3072 4096<br />
user data = window = segment size [byte]<br />
Figure 5 Sparc2 versus Sparc10 segment<br />
processing, SBA-100/2.2.6<br />
159