45 KiB

Исходник Ответственный История

ctsTraffic

ctsTraffic is a highly scalable client/server networking tool giving detailed performance and reliability analytics

If you would like to download the latest build and not have to pull down the source code to build it yourself, you can download them from https://github.com/microsoft/ctsTraffic/tree/master/Releases/2.0.3.5 .

New Visualization Tool!

A great new visualization toolset has been created that can post-process the output files generated by ctsTraffic. At your convenince, please look at https://github.com/microsoft/Network-Performance-Visualization

A Practical Guide

ctsTraffic was a tool initially developed just after Windows 7 shipped to accurately measure how our diverse network deployments scale, as well as assessing its network reliability. Since then we have added a huge number of options to work within an increasingly growing number of deployments. This document reviews the 90% case that most people would likely want to start.

Good-Put

ctsTraffic is deliberately designed and implemented to demonstrate various best-practice guidance we (Winsock) have provided app developers for designing efficient and scalable solutions. It has a "pluggable" model where we have author multiple different IO models -- but the default IO model is what will be most scalable for most network-facing applications.

As our IO models are implemented to model what we want apps and services to build, the resulting performance data is a strong reflection of what one can expect normal apps and services to see in the tested deployment. This throughput measurement of data as seen from the app is commonly referred to as "good-put" (as opposed to "through-put" which is generally measured at the hardware level in raw bits/sec).

A suggested starting point: measuring Good Put

The below set of options (using most default options) is generally a good starting point when measuring good put and reliability. These options will have clients maintain 8 TCP connections with the server, sending 1GB of data per connection. Data will be flowing unidirectionally from the client to the server ('upload' scenarios).

These options will also a good starting point to track the reliability of a network deployment. It provides data across multiple reliability pivots:

Reliably establishing connections over time
Reliably in maintaining connections, sending precisely 1GB of data, with both sides agreeing on the number sent and received
Reliability in data integrity for all bytes received: every buffer received is validated against a specific bit-pattern that we use to catch data corruption

Server	Client
`ctsTraffic.exe`	`ctsTraffic.exe`
`-listen:*`	`-target:<server>`
`-consoleverbosity:1`	`-consoleverbosity:1`
	`-statusfilename:clientstatus.csv`
	`-connectionfilename:clientconnections.csv`

Note: if one needs to measure the other direction, the clients receiving data from servers, one should append -pattern:pull to the above commands on both the client and the server.

We found the above default values to generally be an effective balance when measuring Good Put, balancing the number of connections being established to send and receive data with the number of bytes being sent per connection. We found these values scale very well across many scenarios: down to small devices with slower connections and up to reaching 10Gbit deployments. (Note: once one gets to 10Gb we recommend doubling the number of connections and moving to 1TB of data sent; increasing both again at 40Gb).

Explaining the console output

As a sample run, the below is output from a quick test ran over loopback (client and server were both run on my same machine). Note that the -consoleverbosity: flag controls the type and detail of what it output to the console (setting 0 turns off all output).

C:\Users\kehor\Desktop\2.0.1.7> ctsTraffic.exe -target:localhost -consoleverbosity:1 -statusfilename:clientstatus.csv -connectionfilename:clientconnections.csv

Configured Settings

    Protocol: TCP
    Options: InlineIOCP
    IO function: Iocp (WSASend/WSARecv using IOCP)
    IoPattern: Push \<TCP client send/server recv\>
    PrePostRecvs: 1
    PrePostSends: 1
    Level of verification: Connections & Data
    Port: 4444
    Buffer used for each IO request: 65536 \[0x10000\] bytes
    Total transfer per connection: 1073741824 bytes
    Connecting out to addresses:
           [::1]:4444
           127.0.0.1:4444
    Binding to local addresses for outgoing connections:
           0.0.0.0
           ::
   Connection limit (maximum established connections): 8 \[0x8\]
   Connection throttling rate (maximum pended connection attempts): 1000 \[0x3e8\]
   Total outgoing connections before exit (iterations \* concurrent connections) : 0xffffffffffffffff

Legend:

* TimeSlice - (seconds) cumulative runtime
* Send & Recv Rates - bytes/sec that were transferred within the TimeSlice period
* In-Flight - count of established connections transmitting IO pattern data
* Completed - cumulative count of successfully completed IO patterns
* Network Errors - cumulative count of failed IO patterns due to Winsock errors
* Data Errors - cumulative count of failed IO patterns due to data errors

TimeSlice	SendBps	RecvBps	In-Flight	Completed
0.001	0	0	8	0
5.002	2635357062	124	8	8
10.003	2519263596	171	8	19
15.001	2437002784	202	8	32
20.002	2639655364	171	8	43
25.002	2557516185	218	8	57

Historic Connection Statistics (all connections over the complete lifetime)

SuccessfulConnections [59] NetworkErrors [0] ProtocolErrors [0]

Total Bytes Recv : 5194

Total Bytes Sent : 67358818304

Total Time : 26357 ms.

Configured Settings

The banner under "Configured Settings" shows many of the defaulted options.

Default is to establish TCP connections using OVERLAPPED I/O with IO completion ports managed by the NT threadpool, by default handling inline successful completions. Thus all send and receive requests will loop until an IO request pends, where we will wait for the completion port to notify when it completes, when we'll continue to send and receive.
- IO model is configurable: -IO. Inline IO completions is configurable: -inlineCompletions. Protocol is configurable: -protocol.
Default pattern when sending and receiving is to "Push" data, directionally sending data from the client to the server.
- The pattern to send and receive is configurable: -pattern.
Default is to use 64K buffers for every send and receive request, transferring a total of 1GB of data.
- The buffer size used for each IO request is configurable: -buffer. The total amount of data to transfer over each TCP connection is configurable: -transfer.
Shows all resolved addresses which will be used in a round-robin fashion as connections are made.
- Target IP addresses is configurable using one or more -Target options, specifying a name or IP address servers which to connect.
Shows that will use ephemeral binding (binding to 'any' address of all zeros lets the TCP/IP stack choose the best route to each target address).
- The local addresses to use for outbound connections is configurable: -bind.
Default is to keep 8 concurrent connections established and moving data. Will throttle outgoing connections by only keeping up to 1000 connection attempts in flight at any one time (will above our 8 concurrent connections ).
- The number of connections to establish is configurable: -connections. The connection throttling limit is configurable (though not recommended): -throttleConnections.
Default is to indefinitely continue making connections as individual connections complete -- maintaining 8 connections at all time.
- The total number of connections is configurable: -connections: and -iterations. The total connections is the product of connections * iterations.
- e.g. -connections:100 -iterations:10 will iterate 10 times over 100 connections for a total of 1000 connections, at which point ctsTraffic will gracefully exit.

-consoleverbosity:1

Setting console verbosityto 1 will output an aggregate status at each time slice. The default time slice is every 5 seconds; the time slice is configurable: -statusUpdate. At every 5 seconds, a line will be output communicating the following aggregate information:

TimeSlice: the time window in seconds with millisecond precision, starting from when ctsTraffic was launched (since -statusUpdate can set update frequency in milliseconds)
SendBps: the sent bytes/second [within that specific time slice]
RecvBps: the received bytes/second [within that specific time slice]
Inflight: the [current number] of active TCP connections at this time the slice was recorded
Completed: the [total number] of TCP connections which successfully send and receive all bytes at the time the slice was recorded
NetError: the [total number] of TCP connections which failed due to a Winsock API failing (generally failing to connect, send, or receive) at the time the slice was recorded
DataError: the [total number] of TCP connections which failed due to data validation errors (received too many bytes, received too few bytes before completing the TCP connection, discovered data errors when receiving data) at the time the slice was recorded

This output serves to give the viewer a quick assessment of what is, and has, occurred across the TCP connections that were established. The output functions the same on both the client and the server.

Options for controlling how long a test runs

ctsTraffic has a few options for controlling the amount of time for a run before it exits.

The manual approach is to just hit CTRL-C in the command-shell. ctsTraffic recognizes the key-press and will gracefully exit, ensuring data is accurately flushed to all log files.

The client can control its exit through 2 possible parameter combinations:

-connections and -iterations
- ctsTraffic semantics to track TCP connections are with the sum of these 2 options. If a client needs to eventually go through 1,000 connections but only keeping 100 connections active at any one time, one should say -connections:100 -iterations:10. After 1,000 total connections, ctsTraffic for that client will gracefully exit.
-timeLimit
- ctsTraffic clients can also run for a specified period of time based off the -timeLimit parameter. The parameter expresses the number of milliseconds to run before exiting.

The server can control its exit through just 1 option -- as it is designed to accept any number of connections from any number of clients.

-serverExitLimit
- ctsTraffic servers can gracefully exit automatically after handling the specified number of client connections as set in -serverExitLimit. For example, if one wanted a server to handle 10 connections and exit, one would append -serverExitLimit:10. This ctsTraffic server instance would accept the first 10 inbound connections and complete the specified IO for each of those 10. Any other connections attempted would not be accepted, and after completing those 10 connections the server would gracefully exit.

Explaining the generate log files

In the same sample as above, two log files were created due to the following command line options: -statusfilename:clientstatus.csv -connectionfilename:clientconnections.csv. The csv extension informed ctsTraffic output the files in a comma-separated values format (any other extension would be written as a line of text).

StatusFilename

The status file writes out the same information to a csv as is written to console with the above -consoleverbosity:1 option set. This is useful for later analysis, notably in an application like Excel. Imported into Excel, 25 seconds worth of data would look like this:

TimeSlice	SendBps	RecvBps	In-Flight	Completed
0.001	2	0	8	0
5.002	2635514317	124	8	8
10.003	2519781898	171	8	19
15.001	2437029014	202	8	32
20.002	2639838829	171	8	43
25.002	2557568614	218	8	57

This becomes useful as Excel can quickly give richer views into our data.

Immediately we can look at the last line at Completed (there were 57 successful TCP connections when the last time slice was recorded), NetError and DataError (there were 0 failed TCP connections either through network failures or data errors).

For example, if we wanted to take an average of the SendBps values starting at time slice 5, we would simply specify this in a cell, =AVERAGE(B3:B7). Similarly we can see the min and max with =MIN(B3,B7) and =MAX(B3:B7) respectively. With longer runs and large data sets, this can be notably powerful to understand the overall performance metrics of a run.

Additionally, Excel does quick graphing, which can be some of the most powerful ways to view data. With a larger dataset (with the same above commands specified above), the final graph for SendBps on the client looked like this:

CHART{.chart}

Bits per second was generated by adding a column and telling Excel to multiple SendBps * 8 (the result looks like a consistent ~20 Gbps).

ConnectionFilename

The status file writes out per connection information to a csv (this would be the same as what is written to console with "-consoleverbosity:3" option set). This is useful for later analysis when wanting to look at patterns across a long test run.

Here's an example from the above sample run of the first 10 connections:

{width="6.5in" height="1.3652777777777778in"}

In this log file we can see individual TCP connections recorded.

The TCP tuple information (the local address and port, the remote address and port)
Sent data details (the total # of sent bytes and the SendBps for that one connection)
Received data details (the total # of received bytes and the RecvBps for that one connection)
The total time required for this connection to complete its transfer (TimeMs)
The end result showing success, the error code if a networking error, or the type of data error if a data error was observed
The "ConnectionId" -- a shared GUID from the server and client -- useful when scenarios need to correlate client logs to server logs (even more useful when going through NATs and load balancers where addresses don't always provide unique reference points)

As with the Status file analysis, Excel can give deeper insight into the test run. For example:

One could look at failure patterns: did they occur randomly or in specific groups of time
One could look at which remote addresses had more failures
One could look at patterns in SendBps and RecvBps, correlating with remote servers to see which servers were more performant
One could look at TimeMs to get a broader view over fairness between connections, were some connections starved while some going remarkably faster

CHART{.chart}

As an example, the above 2 charts give a rich view into addressing the "fairness" question across all connections across 5 minutes (over loopback). One could do similar analysis comparing connections across different server addresses to look for issues with servers or groups of servers (e.g. behind a bad routers for example).

A detailed network behavior of the above example

For those inclined, the below explains in more detail the network traffic generated with the above commands.

The default options will result in 8 concurrent TCP connections established from the client to the server.
- The number of connections established can be optionally configured on the client via -connections
The client will send 1GB of data (1073741824 bytes) across each connection using 64KB buffers. The data sent will come from a specifically formatted buffer that is used on both the client and server. This allows for accurate data validation.
- The number of bytes transferred over each connection can be optionally configured, [which must be equivalently set on both the client and server]{.underline}, via -transfer
After sending data the client will wait for a confirmation response from the server that all data was received and verified.
After verifying success or failure of the connection, the connection is closed, and a new connection is established to the server (working to keep 8 concurrent connections alive).
The server accepts all connections from the client. ctsTraffic listening as a server accepts any number of connections from any number of clients.
Once accepted, the server will expect data to be sent from the client and start receiving.
The server, upon every receive completion, will verify the precise bit pattern in the bytes received. Any bytes received that do not match the expected pattern results in immediate failure and the server terminates the connection (immediately calls closesocket).

Scaling

ctsTraffic was deliberation designed to scale: scaling down (it's been used with very small IoT parts to view what good put looks like on a device with very few resources), scaling up (it's been used to large servers with 50 Gbps configurations to look at expected good put), and scaling out (it's been tested with deployments of 10s of thousands of concurrent connections; tested up to 500,000 connections).

We generally recommend scaling to match both the expected nominal deployments and workloads as well as the 90% extreme deployments and workloads. Using the options above to increase the numbers of concurrent connections, ctsTraffic will naturally scale to the resources and network pipes available.

It's useful to note that this scaling comes with the same coding models -- the same code runs which measures small IoT devices without overloading their CPUs as severs with hundreds of cores that run 50Gbps pipes. This all comes with our recommendation: using overlapped I/O with the NT thread-pool and handling inline completions. It's a great demonstration how the Windows OS will scale naturally.

Testing for reliability

We have added features over time which we found greatly helps in measuring the reliability of a networking deployment. Below are examples of combinations of options we have found to be particularly useful in discovering issues in networking components and devices.

Looking for data corruption

While thankfully rare, we have found one method has been particularly successful in discovering data corruption issues in hardware and software stacks. This has found data corruption issues across a variety of vendors and deployments. Interestingly in most cases it was only ctsTraffic and only when ctsTraffic was run in this way was the data corruption observed.

Server	Client
`ctsTraffic.exe`	`ctsTraffic.exe`
`-listen:*`	`-target:<server>`
`-consoleverbosity:1`	`-consoleverbosity:1`
`-pattern:duplex`	`-pattern:duplex`

The unique bit here was running the "full duplex" pattern. This data pattern will send and receive at line rate concurrently: sends posting as quickly as they can post, receives posting as quickly as they can post, all in parallel. This often results in making software work the "hardest" as it must be tracking each TCP stream of data going in both directions, at line rate. "At line rate" was also generally required. With some 40 Gbps network devices we would only discover corruptions when running the duplex pattern at full 40 Gbps bidirectional line rate.

Note that scaling the number of connections and transfer size of each connection as one goes above 1Gbps does also help as it allows more time for each connection. If a network deployment continues to have trouble scaling to line rate, specifying the buffer sizes to 1MB can help (-buffer:1048576).

Randomizing buffer sizes

If one wants to work even harder to find data corruption bugs, one can instruct ctsTraffic to randomize the buffer sizes used for each send and receive request. This will often change the buffering patterns across a networking stack, as TCP segments get created of different sizes which can influence many other TCP factors, such as packet sizes and window sizes.

The default value is 64k for all IO requests on all connections. Randomizing buffer sizes can be done by specifying a range with square brackets. The below is an example where each TCP connection would be randomly choosing a buffer size to use for that connection between 1KB and 1MB.

Server	Client
`ctsTraffic.exe`	`ctsTraffic.exe`
`-listen:*`	`-target:<server>`
`-consoleverbosity:1`	`-consoleverbosity:1`
`-pattern:duplex`	`-pattern:duplex`
`-buffer:[1024,1048576]`	`-buffer:[1024,1048576]`

As noted previously, adjusting numbers of connections and the total transfer size can be useful especially when working on deployments beyond 1Gbps.

Looking for connection establishment issues

If one wants to work even harder to find issue in connection establishment, there are options which can be used to force many more connections to happen over time. The key to doing so is giving a much larger value for the number of connections: -connection [as well as]{.underline} giving a much smaller transfer size: -transfer:. The combination tells ctsTraffic a) maintain a lot of concurrent connections, and b) each connection should be very short-lived.

The result will cycle through a lot of connections very quickly.

Server	Client
`ctsTraffic.exe`	`ctsTraffic.exe`
`-listen:*`	`-target:localhost`
`-consoleverbosity:1`	`-consoleverbosity:1`
`-transfer:64`	`-transfer:64`
	`-connections:100`

These commands when run as a quick test created the below output:

TimeSlice	SendBps	RecvBps	In-Flight	Completed	NetError
5.000	0	0	45	1555	0
10.000	19737	24055	0	3100	0
15.000	19200	23400	0	4600	0
20.000	19200	23400	0	6100	0
25.000	19200	23400	0	7600	0
30.001	19196	23395	0	9100	0
35.000	19203	23404	0	10600	0
40.000	19200	23400	0	12100	0
45.001	19196	23395	0	13600	0
50.000	19203	23404	0	15100	0
55.001	16623	20259	44	16372	154
60.001	9100	11092	33	17110	897
65.000	10037	12224	49	17868	1663

In the output from the run we can see in the Completed column that we were quickly iterating through many thousands of successful connections.

One should also note that at around the 55 second mark we started seeing errors. This is because of a TCP behavior called TIME-WAIT. Because the default behavior for ctsTraffic is for the clients to issue a graceful shutdown at the end of a connection, we create a 4-way FIN to gracefully tear down that TCP connection. While this is a typical way clients and servers terminate connections this can result in the client's tuple (its IP and port) to be temporarily held in a "time-wait" state per RFC. While in these states that port cannot be reused.

This can result in exhausting available ephemeral ports that the client can choose from (even with some recent Windows TCP/IP stack fixes to work harder to potentially reuse some of these ports).

We have options in ctsTraffic which can help to work around this issue: one can tell ctsTraffic how to terminate each successful connection. To avoid entering time-wait, we can tell ctsTraffic to force a RST to shutdown the connection. An RST is a rude/abrupt way to end a connection but is perfectly valid. The command line with this combination would like this:

Server	Client
`ctsTraffic.exe`	`ctsTraffic.exe`
`-listen:*`	`-target:localhost`
`-consoleverbosity:1`	`-consoleverbosity:1`
`-transfer:64`	`-transfer:64`
	`-connections:100`
	`-shutdown:rude`

The -shutdown option (either 'graceful' or 'rude') will instruct the client in how to end their connection (the server will always wait for the client to initiate a closure and therefore never enter time-wait -- something we highly recommend to those building server software). As you see in our simple example instead of seeing failures after about 16,000 connections, we were still creating successful connections after 20,000 connections.

TimeSlice	SendBps	RecvBps	In-Flight	Completed
0.001	0	0	0	0
5.000	20087	24480	34	1566
10.001	19592	23879	0	3100
15.000	19203	23404	0	4600
20.000	19200	23400	0	6100
25.000	19200	23400	0	7600
30.001	19196	23395	0	9100
35.000	19203	23404	0	10600
40.000	19200	23400	0	12100
45.000	19200	23400	0	13600
50.000	19200	23400	0	15100
55.000	19200	23400	0	16600
60.000	19200	23400	0	18100
65.000	19200	23400	0	19600
70.000	19200	23400	0	21100

UDP stream reliability

ctsTraffic measures UDP flows through media streaming semantics -- how most apps (especially client facing apps) use UDP datagrams. In our UDP stream implementation, every datagram is tagged by number and by time. Thus, the client receiving the stream of datagrams from the server can accurately identify every dropped datagram as well as validating the data integrity of each received datagram (the same bit-pattern analysis occurs with UDP as with TCP to check for data corruption).

A suggested starting point: measuring a common UDP stream

It's recommended to start with current stream behaviors -- to replicate and measure those streams over time. To express this in scenario terms, Netflix of often streaming much of its 2160p (4K) content at 15.26 Mbps, though they recommend 25 Mbps availability.

We can accurately measure a deployment's ability to stream a 4K movie at these rates. We will accurately send the specified stream and upon receiving verify the data integrity of all datagrams, track all lost frames (which would translate to lost packets), and track all repeated frames (which can happen with various network topologies).

Server	Client
`ctsTraffic.exe`	`ctsTraffic.exe`
`-listen:*`	`-target:localhost`
`-protocol:udp`	`-protocol:udp`
`-bitspersecond:25000000`	`-bitspersecond:25000000`
`-framerate:60`	`framerate:60`
`-bufferdepth:1`	`-bufferdepth:1`
`-streamlength:60`	`-streamlength:60`
`-consoleverbosity:1`	`-consoleverbosity:1`
	`-connections:1`
	`-iterations:1`
	`-statusfilename:udpclient.csv`
	`-connectionfilename:udpconnection.csv`
	`-jitterfilename:jitter.csv`

These options specify for the client to send a datagram to the server to initiate a "connection" -- where the server will be sending 25Mbps of data across 60 "frames" (datagrams) per second. Buffer depth is how much of a time allowance the client will allow for variance in receiving datagrams. 1 second is generally fine for most simulations.

The result of this test produces 3 log files, 2 similar to the TCP logs and one which tracks jitter by comparing time stamps within the received datagrams.

Explaining the console output

C:\\Users\\kehor\\Desktop\\2.0.1.7\> **ctsTraffic.exe -target:localhost -protocol:udp -bitspersecond:25000000 -framerate:60 -bufferdepth:1 -streamlength:60 -connections:1 -iterations:1 -consoleverbosity:1 -statusfilename:udpclient.csv -connectionfilename:udpconnection.csv -jitterfilename:jitter.csv**

Configured Settings

    Protocol: UDP
    Options: InlineIOCP SO\_RCVBUF(1048576)
    IO function: MediaStream Client
    IoPattern: MediaStream \<UDP controlled stream from server to client\>
    PrePostRecvs: 2
    PrePostSends: 1
    Level of verification: Connections & Data>
    Port: 4444
    Buffer used for each IO request: 52083 \[0xcb73\] bytes
    Total transfer per connection: 187498800 bytes
    UDP Stream BitsPerSecond: 25000000 bits per second
    UDP Stream FrameRate: 60 frames per second
    UDP Stream BufferDepth: 1 seconds
    UDP Stream StreamLength: 60 seconds (3600 frames)
    UDP Stream FrameSize: 52083 bytes
    Connecting out to addresses:
           [::1]:4444
           127.0.0.1:4444
    Binding to local addresses for outgoing connections:
           0.0.0.0
           ::
    Connection limit (maximum established connections): 1 \[0x1\]
    Connection throttling rate (maximum pended connection attempts): 1000 [0x3e8]
    Total outgoing connections before exit (iterations \* concurrent connections) : 1 [0x1]

Legend:

* TimeSlice - (seconds) cumulative runtime
* Streams - count of current number of UDP streams
* Bits/Sec - bits streamed within the TimeSlice period
* Completed Frames - count of frames successfully processed within the TimeSlice
* Dropped Frames - count of frames that were never seen within the TimeSlice
* Repeated Frames - count of frames received multiple times within the TimeSlice
* Stream Errors - count of invalid frames or buffers within the TimeSlice

TimeSlice	Bits/Sec	Streams	Completed
5.000	290	1	240
10.000	24999840	1	300
15.000	24999840	1	300
20.000	24999840	1	300
25.000	25004840	1	300
30.000	24999840	1	300
35.000	24999840	1	300
40.000	24999840	1	300
45.001	24999840	1	300
50.000	25004840	1	300
55.000	24999840	1	300
60.000	25004840	1	300
61.273	327566	0	60

Historic Connection Statistics (all connections over the complete lifetime)

SuccessfulConnections [1] NetworkErrors [0] ProtocolErrors [0]

Total Bytes Recv : 187498800

Total Successful Frames : 3600

Total Dropped Frames : 0

Total Duplicate Frames : 0

Total Error Frames : 0

Total Time : 61273 ms.

The banner under Configured Settings shows default settings with how the streaming parameters were turned into datagram rates.

Default is to establish UDP session using OVERLAPPED I/O with IO completion ports managed by the NT threadpool on the client receiving datagrams (and just blocking sends with an NT threadpool timer on the server), by default handling inline successful completions.
By default, ctsTraffic will configure the client to set the socket option SO_RCVBUF to 1 MB. This pre-allocates 1MB of buffer in Winsock (afd.sys), which is reasonable and suggested for client apps receiving a stream.
- This buffer allows for timing variance with a general purpose OS for IO completions
- This is configurable should one want to emulate other types of apps: -RecvBufValue
Default is to pre-post 2 pended receives (keeping 2 OVERLAPPED receives pended waiting for data). As datagrams do not have in-order guarantees, ctsTraffic accounts for tracking the order of received datagrams (as any app receiving UDP datagrams should).
- This is configurable should one want to push a much greater throughput over a single UDP connection: -PrePostRecvs
Shows the details of how the stream translates into individual datagrams being sent and received.
- Buffer is the buffer size used with each receive request. It generally equates to the calculated FrameSize
- Total transfer is the calculated number of bytes when multiplying the bits/second by the total amount of time the stream will run (and then * 8 to convert to bytes)
- Shows the input BitsPerSecond, FrameRate, BufferDepth, and StreamLength
- Shows the calculated FrameSize -- the number of bytes that will be sent (FrameRate) number of times per second
  - For example, in the above session where we're streaming 25Mbps at 60 frames per second, the server will be sending 52,083 bytes split evenly across 60 times per second.
Shows all resolved addresses which will be used in a round-robin fashion as connections are made.
- Target IP addresses is configurable using one or more -Target options, specifying a name or IP address servers which to connect.
Shows that will use ephemeral binding (binding to 'any' address of all zeros lets the TCP/IP stack choose the best route to each target address).
- The local addresses to use for outbound connections is configurable: -bind.
Default is to keep 1 UDP session ("connection") established and streaming data.
- The number of connections to establish is configurable: -connections.
Default is to indefinitely continue making connections as individual connections complete -- maintaining 8 connections at all time.
- The total number of connections is configurable: -connections: and -iterations:. The total connections is the product of connections * iterations.
- e.g. -connections:100 -iterations:10 will iterate 10 times over 100 connections for a total of 1000 connections, at which point ctsTraffic will gracefully exit.

-consoleverbosity:1 will output an aggregate status at each time slice. The default time slice is every 5 seconds; the time slice is configurable: -statusUpdate. At every 5 seconds, a line will be output communicating the following aggregate information:

TimeSlice: the time window in seconds with millisecond precision, starting from when ctsTraffic was launched (since -statusUpdate can set update frequency in milliseconds)
Bits/Sec: the calculated received bits/sec [within that specific time slice]{.underline}
Streams: the number of concurrent streams running at the time the slice was recorded
Completed: the number of successfully verified frames received [within that specific time slice]{.underline}
Dropped: the number of verified dropped frames [within that specific time slice]{.underline}
Repeated: the number of verified repeated frames [within that specific time slice]{.underline}
Errors: the number of frames with data errors found [within that specific time slice]{.underline}

Explaining the generated log files

In the same sample as above, three log files were created due to the following command line options: "-statusfilename:udpclient.csv -connectionfilename:udpconnection.csv -jitterfilename:jitter.csv". The csv extension informed ctsTraffic output the files in a comma-separated values format (any other extension would be written as a line of text).

StatusFilename

The status file writes out the same information to a csv as is written to console with the above "-consoleverbosity:1" option set. This is useful for later analysis, notably in an application like Excel. Imported into Excel, 25 seconds worth of data would look like this:

TimeSlice	Bits/Sec	Streams	Completed
5	290	1	240
10	24999840	1	300
15	24999840	1	300
20	24994841	1	300
25	25004840	1	300
30	24999840	1	300
35	24999840	1	300
40	24994841	1	300
45.001	24999840	1	300
50	25004840	1	300
55	24994841	1	300
60	24999840	1	300
61.273	327566	0	60

This becomes useful as Excel can quickly give richer views into our data.
With this view when looking at a long stream we can look at patterns of data loss (Dropped > 0), router issues (Repeated > 0), and data integrity issues (Errors > 0).
Additionally, Excel does quick graphing, which can be some of the most powerful ways to view data.
Over loopback the client received a very steady 25 Mbps. As one scales this can certainly change if there are issues with the network supplying a consistent stream of data.

CHART{.chart}

ConnectionFilename

This is similar to the TCP connection view, this time with aggregate data points for each connection.

For a sample, I ran the above 25Mbps run with 10 concurrent UDP streams (-connections:10). Below is the connection output:

{width="6.5in" height="1.3923611111111112in"}

In this log file we can see individual UDP sessions ("connections") recorded.

The UDP tuple information (the local address and port, the remote address and port)
Received bits/second aggregate across the stream
Successfully Completed (received and validated) frames (datagrams)
All verified Dropped datagrams for that stream
All verified Repeated datagrams for that stream
All verified data Errors in received datagrams for that stream
The end Result of that stream
- e.g. were there were Winsock errors which prevented that session from continuing to try to receive datagrams
The "ConnectionId" -- a shared GUID from the server and client -- useful when scenarios need to correlate client logs to server logs (even more useful when going through NATs and load balancers where addresses don't always provide unique reference points)

As with the Status file analysis, Excel can give deeper insight into the test run. For example:

One could look at failure patterns: did they occur randomly or in specific groups of time
One could look at which remote addresses had more failures
One could look at patterns in Bits/Sec, correlating with remote servers to see which servers were more performant

JitterFilename

When making a test run with just a single UDP connection, the client can also track jitter information. This is collected by tracking every individual datagram received and looking at the times stamped on it by the server. Even though the client and server are not time synchronized, the client can still analyze the latency deltas by using the first datagram received as its baseline and calculating gaps. Because the client and server had the same parameters specified, the client know the number of milliseconds that the server would have been waiting between calls to send(). By subtracting the known timer value when the server was waiting between sends it can calculate the time between the actual send and the resulting receive.

As a trivial example, here is jitter being tracked over loopback for the first 30 datagrams received:

{width="6.5in" height="4.928472222222222in"}

The chart shows the sender and receiver QueryPerformanceCounter and QueryPerformanceFrequency values which ctsTraffic stamped in the datagram payload. This leads to being able to calculate "Estimated Received Datagram Time In Flight". You'll note that being over loopback and given optimizations the math resulted in the time in flight being negative .

Streaming over Wi-Fi

As a more interesting example, running the above 25Mbps stream from a small Surface laptop over Wi-Fi to another machine also connected over Wi-Fi shows more diverse data.

The status output now shows more variance in throughput as well as infrequent packet drops. Because this was a shorter run (only 60 seconds) and I wanted to look into greater detail, I set -statusUpdate:500 so I had a twice/second updated view of throughput and packet drops.

Here's a sample of the first 10 seconds:

{width="5.638194444444444in" height="4.595138888888889in"}

You'll notice now that we have data every ½ second (500 ms). We can see we expected to receive 30 individual frames within each ½ second and there were bursts when datagrams were dropped.

Graphing always helps .

CHART{.chart}

This is showing bits/second across the entire 60 seconds of the stream. Because this is Excel I could also quickly do =AVERAGE(), =MIN(), and =MAX() to get a slightly better view into the data:

AVERAGE	24091960
MIN	23333184
MAX	24850735

Just as useful we can look at Jitter data in a more relevant scenario -- here are the first 20 frames:

{width="6.5in" height="3.238888888888889in"}

We can see that variance drifted quite a bit, with larger gaps with a few negative gaps as datagrams arrived in bursts. Graphing this information gives us insightful views into the variance between datagrams received:

CHART{.chart}

With these views we can now see the variance distribution over this 60 second 25 Mbps stream of datagrams.

45 KiB Исходник Ответственный История

ctsTraffic

New Visualization Tool!

A Practical Guide

Good-Put

A suggested starting point: measuring Good Put

Explaining the console output

Configured Settings

-consoleverbosity:1

Options for controlling how long a test runs

Explaining the generate log files

StatusFilename

ConnectionFilename

A detailed network behavior of the above example

Scaling

Testing for reliability

Looking for data corruption

Randomizing buffer sizes

Looking for connection establishment issues

UDP stream reliability

A suggested starting point: measuring a common UDP stream

Explaining the console output

Explaining the generated log files

StatusFilename

ConnectionFilename

JitterFilename

Streaming over Wi-Fi

45 KiB

Исходник Ответственный История