mTCP Performance Notes

Performance Notes

mTCP comes with a program called SpdTest.exe that can be used to measure TCP/IP performance on your machine. The program is simple; it sends or receives a few megabytes of data and times it. The program tries to send and receive data as quickly as it can; no effort is made to manipulate the data because we want to measure the speed of the TCP/IP code and the hardware, nothing else. No shortcuts are taken in the TCP/IP stack - all of the checksums and other things still happen. This type of benchmark is useful because it gives you an upper bound for what to expect for TCP/IP performance.

SpdTest needs to be used against another system that can send and receive data on a socket. I use a much faster Linux system and the netcat command. Command line parameters are:

spdtest <mode> -target <ipaddr> <port> [send options]
spdtest <mode> -listen <port> [send options]

<mode> either -receive (for a receive test) or -send (for a send test). Use -target to have spdtest open the socket connection to a waiting machine or -listen to have spdtest wait for an incoming connection. Port is the port on the target machine when using -target or the local port to listen on when using -listen. The only send option is -mb <n>, where n is the number of megabytes to send.

For example, to execute a send test:

DOS system: spdtest -send -listen 2000 -mb 8
Linux side: time nc 192.168.2.100 2000 -q 0 > bogus.bin

Run the DOS command first so that SpdTest is sitting and waiting for the incoming connection. On the Linux side I use time to measure the wall-time of the entire command and I redirect the output to a file that can be used with the receive test. (On a fast system the cost of doing that is negligible and won't affect your test.) Substitute in the correct IP address for the DOS machine and adjust the port number if you need to. You can also adjust the amount of data to send with the -mb option.

Note: There are at least two versions of netcat out there. I am using the one in the netcat-openbsd package. The other version is found in netcat-traditional and requires you to specify -q 0 to work correctly. Without that option netcat will stay open after reading an EOF from stdin for longer than it should, affecting the test timing. Always specifying -q 0 is safe regardless of which version you have.

To execute a receive test:

DOS system: spdtest -receive -listen 2000
Linux side: time nc 192.168.2.100 2000 -q 0< bogus.bin

Remember bogus.bin? If everything went well during the send test then bogus.bin has 8MB (or whatever you chose), but you can use any or source of data. The idea is to just know how much was sent so that you can do the math.

Tips for running good tests:

Try to run on as close to an isolated network as possible. Excessive broadcast traffic can impact the results, especially on slower systems.
Try to run tests for as long as you can sit still for. On slow systems 4 or 8MB tests are appropriate, while on faster systems you use 64MB.
I run each test five times and pick the three best results, and then take the average of those.
The total time taken is always recorded by the Linux side; some older PCs and Ethernet cards disable interrupts which affects the system clock. Any machine fast enough to run Linux is going to have a better timer.
Always examine the statistics at the end of the test. Checksum errors, retransmits, or sequence number warnings might indicate bad hardware and thus slow the test down.

Here are the raw TCP/IP socket performance numbers for my personal machines:

Personally measured results

Machine Adapter Send Rate
(KB/sec) Recv Rate
(KB/sec) Code Date OS Notes

IBM Convertible 5140 (8088@4.77Mhz)
Xircom PE3-10BT 42.9 22.8 2020-03-07 PC DOS 3.3 4MB transfer, standard parallel port, no hardware IRQ used

IBM PCjr (NEC V20@4.77Mhz) Xircom PE3-10BT 43.4 40.5 2020-03-07 PC DOS 3.3 4MB transfer, bi-directional parallel port, hardware IRQ

IBM PCjr (NEC V20@4.77Mhz) 3Com 3C503 76.7 77.0 2020-03-07 PC DOS 3.3 8MB transfer, 8 bit ISA card on a bus adapter

IBM PCjr (NEC V20@4.77Mhz) WD8003WT 115.3 117.6 2020-03-07 PC DOS 3.3 8MB transfer, 8 bit ISA card on a bus adapter

IBM PC XT (8088@4.77Mhz) NE1000 57.1 59.1 2020-03-07 PC DOS 5.02 4MB transferm 8 bit ISA card

IBM PS/2 Model 25 (8086@8Mhz) 3Com 3C503 190.3 191.9 2020-03-07 PC DOS 6.1 32MB transfer, 8 bit ISA card

Clone 80386DX@40Mhz, 128K L2 cache Davicom DM9008 702.0 1001.8 2020-03-07 PC DOS 5.02 32MB transfer, 16 bit ISA card UM9008 chipset (NE2000 compat)

Compaq 4/33i (80486DX2@66Mhz) Xircom PE3 208.5 202.7
2020-03-07 PC DOS 6.3 8MB transfer, hardware IRQ 7

Compaq 4/33i (80486DX2@66Mhz) 3Com 3C590B 911.8 1005.8 2020-03-07 PC DOS 6.3 32MB transfer, 16 bit ISA card

Pentium 133 LinkSys LNE100 10638.4 10590.3 2020-03-07 PC DOS 5.02 64MB transfer, 100MB/s PCI card

Aptiva E3N (AMD K6-2@300Mhz) IBM 47G7170 452.7 601.5 2020-03-07 Win 98 DOS 32MB transfer, 10MB/s 16 bit ISA card, slower than it should be

IBM NetVista (Pentium 4) RealTek 8139 9831.9 11458.7 2020-03-07 Win 98 DOS 64MB transfer, 100MB/s PCI card

Other reported results

IBM PC 5150 (8088@4.77Mhz) Intel EtherExpress 8/16 67 73 2020-03-07
Reported by Trixter: 8MB transfer, wired Ethernet to wireless network

AT&T PC 6300 (8086@8Mhz) SMC Elite16 (8013 EWC) 163 165 2020-03-07
Reported by Trixter; 8MB transfer, 8KB RAM window on the card

Clone 80386DX@33Mhz, 128K cache 3Com 3C509 1086 886 2020-03-07
Reported by Trixter

Clone 80386DX@33Mhz, 128K cache SMC Elite16 Combo (8013 EWC) 798 1003 2020-03-07
Reported by Trixter; EMC8013 chipset; SMC_WD packet driver which indicates SMC/WD/IBM Ethernet Adapters

Clone 80386DX@33Mhz, 128K cache Intel EtherExpress 8/16 509 616 2020-03-07
Reported by Trixter; 16 bit mode, 32K memory window

Clone 80386DX@33Mhz, 128K cache Intel EtherExpress 8/16 510 577 2020-03-07
Reported by Trixter; 16 bit mode, no memory window

Clone 80386DX@33Mhz, 128K cache Intel EtherExpress 16TP 509 654 2020-03-07
Reported by Trixter

Clone 80386DX@33Mhz, 128K cache Linksys Ether16 642 891 2020-03-07
Reported by Trixter; possibly an NE2000 clone

Clone 80386DX@33Mhz, 128K cache Xircom PE3-10BT 189 104 2020-03-07
Reported by Trixter; standard parallel port, no IRQ (not bi-directional, EPP or ECP

Personally measured results
Machine	Adapter	Send Rate (KB/sec)	Recv Rate (KB/sec)	Code Date	OS	Notes
IBM Convertible 5140 (8088@4.77Mhz)	Xircom PE3-10BT	42.9	22.8	2020-03-07	PC DOS 3.3	4MB transfer, standard parallel port, no hardware IRQ used
IBM PCjr (NEC V20@4.77Mhz)	Xircom PE3-10BT	43.4	40.5	2020-03-07	PC DOS 3.3	4MB transfer, bi-directional parallel port, hardware IRQ
IBM PCjr (NEC V20@4.77Mhz)	3Com 3C503	76.7	77.0	2020-03-07	PC DOS 3.3	8MB transfer, 8 bit ISA card on a bus adapter
IBM PCjr (NEC V20@4.77Mhz)	WD8003WT	115.3	117.6	2020-03-07	PC DOS 3.3	8MB transfer, 8 bit ISA card on a bus adapter
IBM PC XT (8088@4.77Mhz)	NE1000	57.1	59.1	2020-03-07	PC DOS 5.02	4MB transferm 8 bit ISA card
IBM PS/2 Model 25 (8086@8Mhz)	3Com 3C503	190.3	191.9	2020-03-07	PC DOS 6.1	32MB transfer, 8 bit ISA card
Clone 80386DX@40Mhz, 128K L2 cache	Davicom DM9008	702.0	1001.8	2020-03-07	PC DOS 5.02	32MB transfer, 16 bit ISA card UM9008 chipset (NE2000 compat)
Compaq 4/33i (80486DX2@66Mhz)	Xircom PE3	208.5	202.7	2020-03-07	PC DOS 6.3	8MB transfer, hardware IRQ 7
Compaq 4/33i (80486DX2@66Mhz)	3Com 3C590B	911.8	1005.8	2020-03-07	PC DOS 6.3	32MB transfer, 16 bit ISA card
Pentium 133	LinkSys LNE100	10638.4	10590.3	2020-03-07	PC DOS 5.02	64MB transfer, 100MB/s PCI card
Aptiva E3N (AMD K6-2@300Mhz)	IBM 47G7170	452.7	601.5	2020-03-07	Win 98 DOS	32MB transfer, 10MB/s 16 bit ISA card, slower than it should be
IBM NetVista (Pentium 4)	RealTek 8139	9831.9	11458.7	2020-03-07	Win 98 DOS	64MB transfer, 100MB/s PCI card
Other reported results
IBM PC 5150 (8088@4.77Mhz)	Intel EtherExpress 8/16	67	73	2020-03-07		Reported by Trixter: 8MB transfer, wired Ethernet to wireless network
AT&T PC 6300 (8086@8Mhz)	SMC Elite16 (8013 EWC)	163	165	2020-03-07		Reported by Trixter; 8MB transfer, 8KB RAM window on the card
Clone 80386DX@33Mhz, 128K cache	3Com 3C509	1086	886	2020-03-07		Reported by Trixter
Clone 80386DX@33Mhz, 128K cache	SMC Elite16 Combo (8013 EWC)	798	1003	2020-03-07		Reported by Trixter; EMC8013 chipset; SMC_WD packet driver which indicates SMC/WD/IBM Ethernet Adapters
Clone 80386DX@33Mhz, 128K cache	Intel EtherExpress 8/16	509	616	2020-03-07		Reported by Trixter; 16 bit mode, 32K memory window
Clone 80386DX@33Mhz, 128K cache	Intel EtherExpress 8/16	510	577	2020-03-07		Reported by Trixter; 16 bit mode, no memory window
Clone 80386DX@33Mhz, 128K cache	Intel EtherExpress 16TP	509	654	2020-03-07		Reported by Trixter
Clone 80386DX@33Mhz, 128K cache	Linksys Ether16	642	891	2020-03-07		Reported by Trixter; possibly an NE2000 clone
Clone 80386DX@33Mhz, 128K cache	Xircom PE3-10BT	189	104	2020-03-07		Reported by Trixter; standard parallel port, no IRQ (not bi-directional, EPP or ECP

Want to submit a result? If so, take a good set of measurements and forward me the details. I'll post them here!

Some observations:

The Xircom PE3-10BT adapter runs best on a bi-directional parallel port or better. On a standard serial port it will receive data at half the send rate because it can only use the five parallel port control lines for reading from the device. (See the IBM Convertible results for an example.)
The PCjr doesn't have ISA slots. I tested the WD8003, NE1000 and the 3C503 on a special PCjr bus adapter. Otherwise, the same packet drivers and software are used.
The PS/2 machines have bi-directional parallel ports, but use a non-standard port address to flip the direction of the port. The Xircom drivers do not recognize this and thus they have to be run in nibble mode.
The Aptiva with the NE2000 clone should be much faster than it is; it's possible the card is just terrible or the ISA bus is having extra wait states added. (The BIOS does not allow for setting the ISA bus timings.)
The Pentum 133 with the LinkSys card is about fully saturating the Ethernet line. (10638KB/sec works out to 87,000,000+ bits per second.)
The RealTek on the P4 is actually slower than the LinkSys on the Pentium 133 at sending, but it is quite a bit faster when receiving.

Created August 29th, 2009, Last updated December 13th, 2020
(C)opyright Michael B. Brutman, mbbrutman at gmail.com