5. Análisis bivariado
5.5. Ejercicios para el análisis bivariado
Table 6 shows t hat the average Ethernet utilization
of :t s i ngle VAXstation 2000
workstation
running at ypical remote DEC w i ndows application in a cluster
is 0.16 percen t w ith loca l paging, a nd 0. 2'; percent
with remote paging. To verify t he accuracy of the numbers, we measured Et hernet utilization with a
LA1 analyzer for the local paging scenario and
Table 4 Ethernet Traffic: DECnet and Local Area VAXcluster Compone nts
local Paging Remote Paging
Metric (Nu mber) (Percent) (Number) (Percent)
Ethernet packets (total) 1 47 1 1 1 00 1 6902 1 00
D E C n et component 1 07 1 2 73 1 071 2 63
VAXcl u ster component 3999 27 61 90 37
Ethernet bytes (total) 2570772 1 00 4 1 52742 1 00
DEC net component 1 660353 65 1 660353 40
VAXcl uster component 9 1 04 1 2 35 2492404 60
Ethernet Performance of Remote DEC windows Applications
Table 5 Ethernet Traffic: Data and Protocol Com ponents
Local Paging
Remote Paging
Metric
(Number)
Ethernet packets (total) 1 47 1 1
Data component 1 1 558
P rotocol component 31 53
Ethernet bytes (total) 2570765
Data component 1 76 1 1 56
Protocol component 809609
found avnage Ethernet utilization to be 0. 13 per c<::nt, as compared to the 0 . 16 percent predicted by the DECnet!VAXcluster emulator. For remote paging, avnage Ethernet utilization was measured at 0 . 2 3 p<::rcent , as compared to the 0 . 2 5 percent shown with the DECnet!VAXcluster emulator. These comparisons indicat<:: that the protocol emulation , w ith all its in herent assumptions, was reasonably successful in measuring performance impact.
Measurements also were collected from an LAVe located in a software group within Digital . The workgroup had nearly 40 workstations connected to two VAX 8000 disk servers on a single Ethernet segmen t . These were monochrome or color VAXstation 2000 models, equipped with local paging disks and at least 6MB of memory. This was a software development environment where, the activities were primari ly interactive computing with some batch jobs running on the disk servers. All workstations ran DECwi ndows appl ications under the VMS operating system. The most popular DEC net applications were electronic mai l , compu ter conferencing, and other remote DECwindows clients. Some VAX cluster traffic existed , as well as local area transport (LAT) traffic from a number of terminals connected to a terminal server.
On a typical day, the average Ethernet utilization was about 4 percent. This is 0 . 10 percent on average
Table 6 Average Ethernet Util ization of an
LAVe Node Running DECwrite Remotely
Local Remote
Paging
PagingMetric
(Percent)
(Percent)
Ethernet util ization 0. 1 5 0 . 25
DECnet component 0. 1 0 0. 1 0
LAVe component 0.05 0. 1 5
Data component 0. 1 0 0. 1 9
Protocol component 0.05 0.06
Digital Tecbn lcaljournal Vol. 2 No. 3 Summer 1<)90
(Percent)
(Number)
(Percent)
1 00 1 6902 1 00 79 1 2795 76 21 4 1 07 24 1 00 4 1 52757 1 00 69 3 1 88564 77 31 964 1 93 23
per workstation, compared to 0 . 16 percent in our modeled DECwrite environmenr . Although the data in Table 6 shows that the average network use of a single workstation running DECwindows in a clus ter is l ow, a large c luster of workstations can pro duce peaks that are an order of magnitude h igher than the average. For instance, the peak Ethernet utilization observed was 38 percent. Reasons for these peaks include large files being copied over the network or workstations entering or leaving the cluster. A detai led analysis of peaks in Ethernet use in actual LANs was not done hut should be consid ered when applying the resu lts presented in this
paper to a network capacity planning exercise.
Modeling Study
In a previous section, \Ve presented data that char acterized the Ethernet bandwidth requirements of a single workstation running a typical DECwindows application executing remotely. Through the use of a packet-level Ethernet simulation model, this data can be used to predict network performance when many workstations are c lustered on the same Ethernet segment 7 For the DECwrite workload, we drove the simulation model to the point of satura tion of the Ethernet to investigate the theoretical maximum number of workstations that a single Ethernet segment could support. We investigated whether the Ethernet adapter at the disk server(s) could become a bottleneck, and if so, at what load the bottleneck would happen. Finally, by vary ing a few selected input parameters, we used the model to conunent on the performance of different
hypothetical remote DEC windows environments. In an interactive computing environment similar to the one provided by the DECwindows software, it may be desirable to predict the end-to-end or user-perceived response times to perform various functions, such as menu pulldown , w indow deiconification, or mouse movement . Such an anal ysis would capture the effect of network utilization at the user level. To build and validate a model at
DECwindows Program
this level was beyond the scope of our study. How ever, we do include some information on the degra dation in the overall elapsed time of the workload that results from contention at the Ethernet, assum ing that none of the other resources is a bottleneck.
Modeling Methodology
The most important characteristics of Ethernet traffic are the packet size and packet interarrival time d istributions. This model accepts the cumula
tive distributions for packet size and interarrival time that are generated by the DECnet!VAXcluster emulator and uses these distributions to drive t he simulation . The model itself is a closed queuing
model in which each workstation is represented by a transaction that circulates through the model
for the duration of rhe simulation. With each pass through the Ethernet model, the packet size and
arrival rime are assigned to the transaction from rhe distributions that characterize the traffic of the DECwrite workload. The advantage of using the cumula tive distribution technique is that no
assumptions are made about the Ethernet packet size and in rerarrival time distribu tions. This model
allowed us to use separate distributions for di fferent classes of workloads and simulate a user performing
different workload sessions.
The Ethernet simulation model developed for this project captures the functionality and physical principles of the Ethernet. The model was cardully validated against published measurement results
and also against network data collected for rhe
DEC write workload H
Performance Metrics
The following metrics were used i n this study. • Load. The load variable in the simulation is
the number of DEC:windows workstations rhat
are actively executing the remote DEC:w rite
workload. For simplicity, we assumed that the
workstations were a l l similar.
(Note: Ethernet load, packet s ize, and i nr erarrival
time distributions are the input variables to the simulation model. The fol lowi ng are ::t l l outputs from the simu lation.)
• Utilization. Ethernet utilization is computed by d ividing the total number of bits transferred per second by the theoretical maximum
bandwidth of rhe Ethernet ( 10 megabits per sec
ond) for the d uration of the simulation . Unless
90
otherwise specified, this metric refers to average utilization .
• Packet delay. The packet delay consists of the
waiting rime to acquire the channel and the actual transmission time of the packet. Packer delay is usually measured in microseconds as opposed to disk access or processor service
times that are measured in milliseconds. As the load incn:ases, packet delay through the
Ethernet degrades dramatically at a particular
point that we refer to as the kn<:e of th<: curve. • Adapter saturation. The throughput at which the
Ethernet adapter at the d isk serv<:r or computing system saturates is a critical performance metric
in this environment. We consider only on<: adap ter in this study, rhe DEBN I , which is available on the high-end VA X computers. Extending the
analysis ro other adapters is easily done. The sat uration threshold is representcd in t<:rms of the Ethernet util iza t ion level at which the adap ter
saturates for a given mean packet size rath<:r than
the usual packets or me
g
abytes per second .Modeling Results: DECwrite Workload
We first addressed the question of how many workstations actively run ning DECwrite appl ications remotely on a client computing system can be supported on a single Ethernet segmen t .
We assumed that the system o n which these DECwrite client processes would execute had an infi n ite capacity. In other words, content ion for system resources (e.g. , C Pl' , mcmory and d isk J/0) among the DEC:write clients was not incorpo rated in the model . Because any such contention
woul d reduce network traffic intensity, we pre sented an upper-bound or worst-cas<.: analysis. We also assumed that there was no communication
among the workstations, which would he rrue when all applications were run rc:mo tely. The sim ulation was run for both local paging and remote paging scenarios.
Figure () shows that the average: Ethernet utiliza tion curves increase with load and then !<:vel off at 600 workstations (60 percent utilization) with local
paging and 400 workstations (69 pcrcmt miliza tion) with remote paging. The DEBNI threshold in Figure 6 a lso shows that the Ethernet adapter would saturate at :)')0 workstations with local paging and at 300 workstations w irh remote paging. In Fig ure 7, t he average packet delay cu rves indicate that
the knee in the curve is at a much lower load of :)00
Ethernet Performance of Remote DEC windows Applications
workstations with local paging and 200 work
stations with remote paging. Also indicated in this figure are the points at which network congestion
causes the elapsed time for the workload to degrade
by 10 percent and 100 percent.
We used the point at which packet delay started to degrade, in Figure 7, as the limiting factor. With this criterion, the theoretical size of an LAVe system in a typical remote DECwindows environment would be about 300 active workstations, assuming all of the satellites have local paging disks and steady-state operation. Further, the d isk server and DECwrite clients might need to be distributed over multiple systems to obtain the required processing power especially if lower capacity Ethernet adap ters are being used . (Note: These are average num bers and the user-perceived response time might degrade if large amounts of data are transferred often or if many nodes frequently transition in and out of the c luster. )