The purpose of our first experiment is to quantitatively characterize the limitations of the simple request generation scheme described in Section 3. We ran an increasing number of client processes distributed across 4 client machines. Each client tries to establish a HTTP connection to the server, sends a request, receives the response and then repeats the cycle. Each HTTP request is for a single file of size 1294 bytes. We measured the request rate (incoming SYNs/second) at the server.
In a similar test we ran 12 S-Clients distributed across the 4 client machines with an increasing number of descriptors per S-Client and measured the request rate seen at the server. Each S-Client had the connection establishment timeout period T set to 500ms. The same file was requested as in the case of the simple clients.
Figure 5: Request rate versus number of descriptors
Figure 6: Request rate versus number of clients
Figure 5 plots the total connection request rate seen by the server versus the total number of client processes for the simple client test. Figure 6 plots the same metric for the S-Client test, but with the total number of descriptors in the S-Clients on the x-axis.
For the reasons discussed earlier, the simple scheme generates no more than about 130 requests per second (which is the capacity of our server for this request size). At this point, the server can accept connections at exactly the rate at which they are generated. As we add more clients, the queue length at the accept queue of the server's listen socket increases and the request rate remains nearly constant at the capacity of the server.
With S-Clients, the request rate increases linearly with the total number of descriptors being used for establishing connections by the client processes. To highlight the difference in behavior of the two schemes in this figure, we do not show the full curve for S-Clients. The complete curve shows a linear increase in request rate all the way up to 2065 requests per second with our setup of four client machines. Beyond this point, client capacity resource limitations set in and the request rate ceases to increase. More client machines are needed to achieve higher rates. Thus we see that S-Clients enable the generation of request loads that greatly exceed the capacity of the server. The generated load also scales very well with the number of descriptors being used.
