Latency Performance Tests

Twoway Latency

For this test, the client first makes one call to the server in order to establish a connection. The time needed for this initial call is not taken into account, as the time for connection establishment is not an aspect of latency. After the initial call, the client sends 100,000 twoway invocations to the server, without any parameters or return values. The test calculates the average time for each invocation.

Twoway Latency

	Ice	TAO	Difference
Linux	119.20 µs	291.50 µs	145%
Windows XP	123.30 µs	218.73 µs	77%

Ice is about 2.5 times faster than TAO under Linux, and about 1.8 times faster under Windows XP. Note that these results reflect quality of implementation, as Ice has no signifcant design advantage over CORBA with respect to latency.

Twoway Latency with AMI

Both Ice and CORBA support Asynchronous Message Invocation (AMI). This is the same test as the twoway latency test, but using asynchronous invocations instead of synchronous twoway invocations.

Twoway Latency with AMI

	Ice	TAO	Difference
Linux	132.47 µs	423.37 µs	220%
Windows XP	132.52 µs	312.64 µs	136%

For asynchronous invocation, Ice is about 3.2 times faster than TAO under Linux, and about 2.4 times faster under Windows XP.

Oneway Latency

This test was performed under the same conditions as the twoway latency test, except that oneway invocations were used. After the last of the 100,000 oneway invocations, the test makes one final twoway invocation. This is necessary because the TCP/IP protocol stack can buffer data and return control back to the caller before the buffered data has been transmitted; because the final twoway call requires a reply from the server, it ensures that all previously buffered data has actually been sent.

Oneway Latency

	Ice	TAO	Difference
Linux	31.32 µs	132.14 µs	322%
Windows XP	42.77 µs	78.47 µs	83%

Ice is about 1.8 times faster than TAO under Windows XP, and 4.2 times faster under Linux.

Batched Oneway Latency

Batched performance tests require special consideration. Batched oneway means that several oneway invocations are buffered by the Ice core, which are then sent together in a single request across the network. For this particular test, we used 1,000 batches of 100 oneway invocations. CORBA does not support batched oneways, so we used 100,000 regular oneways.

Batched Oneway Latency

	Ice	TAO	Difference
Linux	7.73 µs	132.14 µs	1609%
Windows XP	16.78 µs	78.47 µs	368%

While the results for non-batched oneway latency (just as the results for twoway latency) can be attributed to quality of implementation, the performance advantage of batched oneway invocations is due to design advantages of Ice over CORBA.

As can be seen from the results, the ability to send batched oneway invocations can yield large performance gains. For example, under Linux, using batched oneways with Ice, the test runs 17 times faster than using regular oneways with TAO.

 

(Return to the Performance Test Index)

Oneway vs. Twoway

Latency is an important aspect of any middleware: it determines the overhead for a remote operation call that has no (or only a very small) payload. The lower the latency, the more invocations per second a client can send to a server. There are two kinds of latency measurements. The first is the latency for a full round trip, meaning the time it takes for the client to send a request, plus the time it takes for the server to return a response back to the client. Such requests are called twoway requests. In contrast, there are also oneway requests, that is, requests for which the server does not send any response. The latency for oneway requests is usually lower because the client does not need to wait for a reply to each request before it can send another request.