Benchmark for ACSflix commit ce8e987501

This benchmark is based on the same test taken by ZMQ, which consisted in:

sending 1M messages over the network as fast as possible. The test was run for different messages size ( 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096 and 8192 bytes). The whole test was performed 3 times.

With that in mind, we followed ZMQ Instructions on How To Benchmark

Latency/Throughput Test

These tests will allow us to define a baseline behaviour

Idle systems, MTU 1500, with different message sizes, for 1 million roundtrips, from A to B:

Message Size (Bytes)	2	4	8	16	32	64	128	256	512	1024
Latency (us)	128.261	130.060	131.086	130.633	132.878	136.549	138.214	139.322	142.327	156.682
Throughput (msg/s)	5.247.773	5.154.931	4.935.712	4.584.653	2.919.801	1.497.671	748.864	365.747	186.768	94.288
Throughput (Mbit/s)	83.964	164.958	315.886	586.836	747.469	766.808	766.837	749.050	765.002	772.407

Message Size (Bytes)	2048	4096	8192
Latency (us)	145.435	264.269	289.649
Throughput (msg/s)	47.470	23.771	11.812
Throughput (Mbit/s)	777.748	778.928	774.111

ACSflix

The logic in ACSflix is pretty simple:

We use PUSH/PULL sockets to send the data.
And a PUB/SUB channel to notify the worker of the End of Transmission.

Each of the Push/Pull pair of sockets is referred as a Flow, which can be grouped in a Stream.

The minimum buffer size is 2 Bytes.
Testings are performed over OSes with firewalls.
An equivalent file size has been calculated with: (1 Million messages * X Bytes) / 1024^2
In ACSflix, using -b flag, buffer size can be set.
Files used are images available at [http://repo.csrg.cl/images]-
Raw Spreadsheat with results here

1 Flow, 1 Stream

File Size [MB] - Buffer Size [B]	1.91 (2)	3.82 (4)	7.64 (8)	15.28 (16)	30.56 (32)	61.12 (64)	122.24 (128)	244.48 (256)
Message Count	1003567	1186428	1066113	1055297	1055155	1012879	995589	1007966
Throughput (msg/s)	534282	498097	526471	528386	531906	482857	490666	389603
Throughput (Mbit/s)	8.549	15.939	33.694	67.633	136.168	247.223	502.442	797.907

File Size (MB)	1956 (2048)	3912 (4096)	7824 (8192)
Message Count
Throughput (msg/s)
Throughput (Mbit/s)

3 Flows, 1 Stream

3 Flows, 3 Streams

HW Specifications

Computer A (Zoidberg): Dell Precision WorkStation T3400, 4GiB RAM DDR2 @ 800 MHz, Core 2 Duo E4600 @ 2.4 GHz, NetXtreme BCM5754 Gigabit
Computer B (acscb): HP Compaq 6200, 6GiB RAM DDR3 @ 1067 MHz, Core i5-2400@ 3.10GHz, 82579LM Gigabit Network
Switch: TP Link TL-SG1008 Gigabit

The computer’s SO is Centos 7.3, kernel 3.10.0-514.10.2

Sources and Useful Links

ZMQ oficial 10Gb Ethernet Tests
[http://nikolaveber.blogspot.cl/2011/04/if-you-are-planning-large-or-not-even.html]
ZMQ Benchmaks
Pub/sub reports 15 us extra interval and ZMQ_NOBLOCK lows the extra time to just 10 us