Fwd: DM6446 Ethernet MAC performance
From: Pablo Bitton pablo.bit...@gmail.com Date: Mon, Mar 2, 2009 at 11:52 AM Subject: Re: DM6446 Ethernet MAC performance To: Subrahmanya, Chaithrika chaithr...@ti.com Hi, I've found that the issue below can be demonstrated by compiling the following two revisions: http://git.kernel.org/?p=linux/kernel/git/khilman/linux-davinci.git;a=commit;h=73af72ed66644dc3703b0bfd74285f30f1fe408f - before PHY refactoring http://git.kernel.org/?p=linux/kernel/git/khilman/linux-davinci.git;a=commit;h=b2cbe2f96aa420efcd312a1de8f7fd8e511b2a55 - after PHY refactoring Build commands used: $ make distclean $ make davinci_dm644x_evm_defconfig $ make uImage The kernels above were loaded to the same EVM, using the same network setup - and had the following performace: 73af72ed66644dc3703b0bfd74285f30f1fe408f - 72Mbit/s (upload) b2cbe2f96aa420efcd312a1de8f7fd8e511b2a55 - 21Mbit/s (upload) This experiment was repeated several times, so probably network congestion is not the reason for performance degradation seen. Regards, Pablo. On Wed, Feb 25, 2009 at 7:44 PM, Pablo Bitton pablo.bit...@gmail.com wrote: My host is Ubuntu 8.04, running over VMware, connected to DVEVM by 2 switches. I agree that this setup may cause network congestion. However, I've ran iperf over DVEVM and the host several times - and the results were pretty much the same. The only thing that made the throughput to change dramatically was switching back and forth between 2.6.29-rc4 and 2.6.28-rc6. How should I debug this issue? Are there any recommended profiling tools for kernel's network performance? Thanks in advance. On Tue, Feb 24, 2009 at 1:39 PM, Subrahmanya, Chaithrika chaithr...@ti.com wrote: Hi to all, I've tested Linux kernel 2.6.29-rc4-davinci1 (http://git.kernel.org/?p=linux/kernel/git/khilman/linux- davinci.git;a=commit;h=582ad6bcceb7c8f0c828e71e810fd40fed7af6c3) for Ethernet performance. Kernel was compiled was done by CodeSourcery 2008q3 (gcc 4.3.2) using davinci_evm_dm644x_defconfig. After booting, iperf v2.0.4 was run on the target (DVEVM board) and on the host PC (Ubuntu 8.04) for ten seconds (using default settings): target - target (over localhost): 190Mbit/s host (connecting to) - target: 31Mbit/s target (connecting to) - host: 71Mbit/s Do the values above make sense? Can anyone confirm them? The PHY layer should not cause any decrease in the performance as it is used only to access the PHY using the MII. Please provide some more details like the connection setup used- whether it was on a network or back to back. The network dynamics will alter the performance. Regards, Chaithrika Thanks in advance. P.S. When running the same test on 2.6.28-rc6 (from http://git.kernel.org/?p=linux/kernel/git/khilman/linux-davinci.git), before PHY was refactored out from davinci_emac, the results were better: host (connecting to) - target: 73Mbit/s target (connecting to) - host: 57Mbit/s ___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source ___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source
RE: DM6446 Ethernet MAC performance
From: Pablo Bitton pablo.bit...@gmail.com Date: Mon, Mar 2, 2009 at 11:52 AM Subject: Re: DM6446 Ethernet MAC performance To: Subrahmanya, Chaithrika chaithr...@ti.com Hi, I've found that the issue below can be demonstrated by compiling the following two revisions: http://git.kernel.org/?p=linux/kernel/git/khilman/linux- davinci.git;a=commit;h=73af72ed66644dc3703b0bfd74285f30f1fe408f - before PHY refactoring http://git.kernel.org/?p=linux/kernel/git/khilman/linux- davinci.git;a=commit;h=b2cbe2f96aa420efcd312a1de8f7fd8e511b2a55 - after PHY refactoring Build commands used: $ make distclean $ make davinci_dm644x_evm_defconfig $ make uImage The kernels above were loaded to the same EVM, using the same network setup - and had the following performace: 73af72ed66644dc3703b0bfd74285f30f1fe408f - 72Mbit/s (upload) b2cbe2f96aa420efcd312a1de8f7fd8e511b2a55 - 21Mbit/s (upload) This experiment was repeated several times, so probably network congestion is not the reason for performance degradation seen. I am looking into this issue, will get back to you with my findings. I have not observed any large difference in performance numbers when I tested it earlier (for UDP). In fact there was some improvement for some cases. Thanks, Chaithrika Regards, Pablo. On Wed, Feb 25, 2009 at 7:44 PM, Pablo Bitton pablo.bit...@gmail.com wrote: My host is Ubuntu 8.04, running over VMware, connected to DVEVM by 2 switches. I agree that this setup may cause network congestion. However, I've ran iperf over DVEVM and the host several times - and the results were pretty much the same. The only thing that made the throughput to change dramatically was switching back and forth between 2.6.29-rc4 and 2.6.28-rc6. How should I debug this issue? Are there any recommended profiling tools for kernel's network performance? Thanks in advance. On Tue, Feb 24, 2009 at 1:39 PM, Subrahmanya, Chaithrika chaithr...@ti.com wrote: Hi to all, I've tested Linux kernel 2.6.29-rc4-davinci1 (http://git.kernel.org/?p=linux/kernel/git/khilman/linux- davinci.git;a=commit;h=582ad6bcceb7c8f0c828e71e810fd40fed7af6c3) for Ethernet performance. Kernel was compiled was done by CodeSourcery 2008q3 (gcc 4.3.2) using davinci_evm_dm644x_defconfig. After booting, iperf v2.0.4 was run on the target (DVEVM board) and on the host PC (Ubuntu 8.04) for ten seconds (using default settings): target - target (over localhost): 190Mbit/s host (connecting to) - target: 31Mbit/s target (connecting to) - host: 71Mbit/s Do the values above make sense? Can anyone confirm them? The PHY layer should not cause any decrease in the performance as it is used only to access the PHY using the MII. Please provide some more details like the connection setup used- whether it was on a network or back to back. The network dynamics will alter the performance. Regards, Chaithrika Thanks in advance. P.S. When running the same test on 2.6.28-rc6 (from http://git.kernel.org/?p=linux/kernel/git/khilman/linux- davinci.git), before PHY was refactored out from davinci_emac, the results were better: host (connecting to) - target: 73Mbit/s target (connecting to) - host: 57Mbit/s ___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open- source ___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source
Re: DM6446 Ethernet MAC performance
My host is Ubuntu 8.04, running over VMware, connected to DVEVM by 2 switches. I agree that this setup may cause network congestion. However, I've ran iperf over DVEVM and the host several times - and the results were pretty much the same. The only thing that made the throughput to change dramatically was switching back and forth between 2.6.29-rc4 and 2.6.28-rc6. How should I debug this issue? Are there any recommended profiling tools for kernel's network performance? Thanks in advance. On Tue, Feb 24, 2009 at 1:39 PM, Subrahmanya, Chaithrika chaithr...@ti.com wrote: Hi to all, I've tested Linux kernel 2.6.29-rc4-davinci1 (http://git.kernel.org/?p=linux/kernel/git/khilman/linux- davinci.git;a=commit;h=582ad6bcceb7c8f0c828e71e810fd40fed7af6c3) for Ethernet performance. Kernel was compiled was done by CodeSourcery 2008q3 (gcc 4.3.2) using davinci_evm_dm644x_defconfig. After booting, iperf v2.0.4 was run on the target (DVEVM board) and on the host PC (Ubuntu 8.04) for ten seconds (using default settings): target - target (over localhost): 190Mbit/s host (connecting to) - target: 31Mbit/s target (connecting to) - host: 71Mbit/s Do the values above make sense? Can anyone confirm them? The PHY layer should not cause any decrease in the performance as it is used only to access the PHY using the MII. Please provide some more details like the connection setup used- whether it was on a network or back to back. The network dynamics will alter the performance. Regards, Chaithrika Thanks in advance. P.S. When running the same test on 2.6.28-rc6 (from http://git.kernel.org/?p=linux/kernel/git/khilman/linux-davinci.git), before PHY was refactored out from davinci_emac, the results were better: host (connecting to) - target: 73Mbit/s target (connecting to) - host: 57Mbit/s ___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source ___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source
RE: DM6446 Ethernet MAC performance
Hi to all, I've tested Linux kernel 2.6.29-rc4-davinci1 (http://git.kernel.org/?p=linux/kernel/git/khilman/linux- davinci.git;a=commit;h=582ad6bcceb7c8f0c828e71e810fd40fed7af6c3) for Ethernet performance. Kernel was compiled was done by CodeSourcery 2008q3 (gcc 4.3.2) using davinci_evm_dm644x_defconfig. After booting, iperf v2.0.4 was run on the target (DVEVM board) and on the host PC (Ubuntu 8.04) for ten seconds (using default settings): target - target (over localhost): 190Mbit/s host (connecting to) - target: 31Mbit/s target (connecting to) - host: 71Mbit/s Do the values above make sense? Can anyone confirm them? The PHY layer should not cause any decrease in the performance as it is used only to access the PHY using the MII. Please provide some more details like the connection setup used- whether it was on a network or back to back. The network dynamics will alter the performance. Regards, Chaithrika Thanks in advance. P.S. When running the same test on 2.6.28-rc6 (from http://git.kernel.org/?p=linux/kernel/git/khilman/linux-davinci.git), before PHY was refactored out from davinci_emac, the results were better: host (connecting to) - target: 73Mbit/s target (connecting to) - host: 57Mbit/s ___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source
DM6446 Ethernet MAC performance
Hi to all, I've tested Linux kernel 2.6.29-rc4-davinci1 (http://git.kernel.org/?p=linux/kernel/git/khilman/linux-davinci.git;a=commit;h=582ad6bcceb7c8f0c828e71e810fd40fed7af6c3) for Ethernet performance. Kernel was compiled was done by CodeSourcery 2008q3 (gcc 4.3.2) using davinci_evm_dm644x_defconfig. After booting, iperf v2.0.4 was run on the target (DVEVM board) and on the host PC (Ubuntu 8.04) for ten seconds (using default settings): target - target (over localhost): 190Mbit/s host (connecting to) - target: 31Mbit/s target (connecting to) - host: 71Mbit/s Do the values above make sense? Can anyone confirm them? Thanks in advance. P.S. When running the same test on 2.6.28-rc6 (from http://git.kernel.org/?p=linux/kernel/git/khilman/linux-davinci.git), before PHY was refactored out from davinci_emac, the results were better: host (connecting to) - target: 73Mbit/s target (connecting to) - host: 57Mbit/s ___ Davinci-linux-open-source mailing list Davinci-linux-open-source@linux.davincidsp.com http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source