FUNDAMENTALES EN EL SISTEMA INSTITUCIONAL Y JURÍDICO DE CUBA.
L A PENA DE MUERTE POR CAUSAS POLÍTICAS
V. L A CONCENTRACIÓN DEL PODER DE LOS ÓRGANOS DEL E STADO
CACHE H I T
tially reads data, the seek time is small because the data within a file is assumed to be contiguous. When severa l processes access data on the same disk, the seek time is higher since the head must move about the disk to access different fi les . We measured the average server CPU time per operation and the indi vidual response times of the server and client.
From our measurements, we noticed that a cache miss almost always resulted in an extra 20 millisec onds (ms) in the response time. To explain this, we considered that the RA-series disks rotate at 3600 rpm (16.67 ms per revolution) , and the RA81 , in panicu lar, stores approximately three 16-block buffers per track. In the case of a single contiguous file being read 16 blocks at a time, each read eventually has to be sync hronized with the revolution of the disk. Moreover, after every three reads, a single-track seek is necessary. This explanation coincides with the 20-ms delay that was measured . When several files are being simultaneously accessed on the disk, the head is required to seek the next request. (The aver age RA81 seek time is 28 ms .) This seek could over lap with the rotational delay and result in a total access time of approximately 45 ms.
Analytical Results Mean value analysis provides a way of examining the delays that the average request is expected to encounter. This analysis does not involve itself in the distributions of the individual delays, which may frequently be theoretically impossible to derive. Using this approach, we attempted to formulate an equation for the delay experienced by a single request .
In a system that has Ndfs client processes, an aver age server CPU time of T.pu, and an average of N,p,
requests waiting for the CPU, the time in and wait ing for the server CPU, T,..c, is
Similarly, the time spent in and waiting for the
disk subsystem, Tu·d, is
T,." = N" X T" + T"
PROCESSOR
CLI ENT THINK TIME
Figure 2 Model of the DFS System
The total round trip time for the average request is given by
Tc, + Twc + Pm X Tux� .
Therefore, if we look at the system from the point of view of one of the clients, we expect Ncpu and Net to be
N,pu = (Netfs - 1 ) X T wc/( Tc, + T.vc + p, X Twct)
Net = (Netfs - 1 ) X Twa/ ( T., + Tux + Pm X T.vc�)
where p, is the probability that a cache miss occurs. This result could also be obtained from Litt le's Law if we consider that the effective traffic seen by the current request is that imposed by the other request.3
N = X T
where A = (Netfs - 1 )/ ( T., + T we + p, X Twet) , the arrival rate of the other requests to the resource, and
T, the service time of the resource.
Likewise, the time for the server to satisfy the request wou ld thus be
Tsen"-"' = TW<- + p, X Twa
These equations are of special interest when we consider the following cases:
Case 1: p, = 0, Nafs = 1. I n case 1 a single client repeatedly accesses a lim ited amount of data (cache hits) . The equations above reduce to
Case 2: Pm = 1 , Netfs = 1 . In case 2 a single client
always requests new data (cache misses) . The equa tions above now reduce to
Case 3: Pm = 0, Nctfs > > 1 . In case 3 several clients bottleneck the server, and the cache is always hit. The time in the server is now given by
This resu lt can also be obtained from a fluid approximation of the syste m , with the assumption that the CPU is always fully util ized and every request is either in the queue for the CPU or in service. Simulation Results and Validation To give our selves a greater amount of flexibil ity and detail, we wrote a simulation model of the DFS system, shown in Figure 2 . A modification in the model accounted for post-I/O processing and disk service times on shared disks. To validate the model, a program that sequentially reads data from a large file ran from one to six processes, and the throughput gained was measured. This measurement was compared to the throughput predicted by the simulation model for
82
two cases: when all different files were resident on a single disk, and when the data was equally shared among four disks . As Figure 3 indicates, the simu lated throughput was within 10 percent of that mea sured for al l cases.
For a single user the throughput of the system is 1 .05 megabits (Mb) per second . For several users, the maximum throughput measured with a MicroVAX II server was found to be 2.6Mb per second. These resu lts were also confirmed using a program that consistently reads data directly from the cache . Since the server CPU was being ful ly uti l ized at this po int, increasing the capacity of any of the other components wou ld not have improved performance. However, the easy-to-use technique of bottleneck analysis described below allowed us to arrive at the same resu lt with l ittle or no effort.
Bottleneck Analysis When configuring large sys tems in which the type of traffic is known, bottle neck analysis gives a good est imate of the system 's maximum capacity. Associating a service time with each component and assuming complete asynchro nic ity of the operations, we used this method to predict the throughput for different configurations. This ana lysis assumes that the slowest component, i . e . , the bottleneck, invariably determines the maxi mum throughput of the whole system in the same way that fewer lanes on a highway determines how many drivers using that road get to work on time.
0 3.50 z 0 3.00 (.) UJ (f) a: 2.50 UJ Cl. .0 2.00 � 1- 1 .50 ::::> Cl. I 1 .00 (.') :J 0 a: 0.50 I 1- 0 0 2 4 NUMBER OF U S E R S KEY: o S I M U LAT ED R EA D, 4 DISKS
o MEASURED READ. 4 DISKS b. S I M U LATED READ, 1 DISK + MEASURED READ, 1 DISK
6
Figure 3 Simulated versus Measured Results of ljO Processing
8
Measurement and Analysis Techniques for DEC net Products
Ta ble 1 Service Times of DFS Components Service
Resource Time
Server CPU , M icroVAX II 4.9 + 1 .26 x blocks ms
Server CPU, VAX 8700 1 .4 + 0.28 x blocks ms Disk time for contiguous data 20 ms
Disk time for noncontig uous 44 ms
data
Effective network service time 1 .08 x blocks ms
Client CPU, MicroVAX I I 6.2 + 1 .26 x blocks ms
(continuous read)
Client CPU, VAX 8700 1 .4 + 0.28 x blocks ms
(continuous read)
Ta ble 1 lists the service time for each component of a DFS system . We assume that the most frequent operation is the read primitive.
The maximum throughput that a system of several clients and a single server can thus process is given by
T = 1jmax ( Tcp11 , p, X Td/D, T.u:r , Tel/C) where D is the nu mber of DFS disks serviced by the CPU, T""' is the effective time to transfer the data across the network, and C is the number of c lients
that use the DFS server.
In the case of users on C -client CPUs accessing noncontiguous data on a MicroVAX II server with an
RA81 disk, an average read size of 6 blocks, and Pm equals 0.25, the throughput is thus
1jmax (4 .9 + 1 . 26 X 6, 0.25 X 4 4 , 1 .08 X 6, (6.2 + 1 . 26 X 6)/C)
That is, for 73 operations per second, n equals 1 ; and for 80 operations per second, C is greater than 1 . Sample Workload Characterization
Approaching the problem from a different perspec tive, we examined installed DFS systems to obtain information about real-life workloads. The VAXcluster system ca l led Server 1 is a DFS server acting primar ily as a database for product save-sets, documenta tion, and other such reference data. The cluster cal led Server 2 is simi larly used, though with a greater developmental and engi neering bias . The statistics obtained are shown in Table 2 .
Considering a file read sequentia lly n blocks at a time, the cache miss ratio is nj16. Thus an average read size of 3 . 4 b locks wou ld imply a cache miss ratio of 21 percent. This coincides remarkably with the Server 1 system's measured cache hit ratio of 23 percent . For the Server 2 cluster, the predicted and measured values are 84 percent and 86 percent,
Digital Teclmicaljournal No. 9 june 1989
Table 2 Workload Statistics for Two Servers Server 1 Server 2 Cluster Cluster Approximate number 20,591 K 81 6K of operations Probability of persona 0.2% 0.5% cache miss Probability of data 23.0% 86.0%
cache miss (for read)
Average read size 3.4 blocks 1 3.5 blocks (approximate)
Average write size -1 6.0 blocks -1 6.0 blocks
(approximate)
Readfwrite ratio 20: 1 1 1 : 1
respectively. We thus concluded that in both systems, files are very seldom already in the cache when ini tial ly requested by a user, and the major benefit of caching is that small access sizes do not necessarily result in disk accesses .
On the DFS server, the persona cache keeps track of user authorization i nformation. The persona cache and its impl ications have been omitted in this paper precisely because this cache is so effective that a cache hit almost always occurs. 1