Re: How to measure the RAM read/write performance
) [2.385986] Stack: (0xe982bf98 to 0xe982c000) [2.390319] bf80: 3ffd a27e [2.398498] bfa0: c0037b4c c0022e94 c003f3fc e9814a80 00373231 [2.406646] bfc0: 0236 c0037b4c c003818c c00461b4 0013 [2.414825] bfe0: c0008374 c0008300 c00461b4 c00461b4 [2.423004] [c0022ee0] (sand_misc_init+0x4c/0xac) from [c003f3fc] (do_one_initcall+0xd0/0x1a4) [2.431915] [c003f3fc] (do_one_initcall+0xd0/0x1a4) from [c0008374] (kernel_init+0x74/0x118) [2.440673] [c0008374] (kernel_init+0x74/0x118) from [c00461b4] (kernel_thread_exit+0x0/0x8) [2.449462] Code: e58d4000 e3033fff ea07 e59d2000 (e7952002) [2.455566] ---[ end trace f76f3c76dcb9b9ef ]--- [2.460144] Kernel panic - not syncing: Attempted to kill init! [2.466064] [c004ad10] (unwind_backtrace+0x0/0x12c) from [c064af70] (panic+0x90/0x1bc) [2.474304] [c064af70] (panic+0x90/0x1bc) from [c00eb8dc] (do_exit+0xb8/0x734) [2.481842] [c00eb8dc] (do_exit+0xb8/0x734) from [c0048f8c] (die+0x208/0x23c) [2.489318] [c0048f8c] (die+0x208/0x23c) from [c004e158] (__do_kernel_fault+0x64/0x84) [2.497558] [c004e158] (__do_kernel_fault+0x64/0x84) from [c004e3e0] (do_page_fault+0x268/0x288) [2.506683] [c004e3e0] (do_page_fault+0x268/0x288) from [c003f270] (do_DataAbort+0x34/0x94) [2.515350] [c003f270] (do_DataAbort+0x34/0x94) from [c0044dac] (__dabt_svc+0x4c/0x60) [2.523590] Exception stack(0xe982bf50 to 0xe982bf98) [2.528625] bf40: 0063 c07ea2cf 3ffd 3fff [2.536804] bf60: ea828000 c00461b4 [2.544952] bf80: e982bf98 c0022ed0 c0022ee0 8013 [2.551544] [c0044dac] (__dabt_svc+0x4c/0x60) from [c0022ee0] (sand_misc_init+0x4c/0xac) [2.559967] [c0022ee0] (sand_misc_init+0x4c/0xac) from [c003f3fc] (do_one_initcall+0xd0/0x1a4) [2.568908] [c003f3fc] (do_one_initcall+0xd0/0x1a4) from [c0008374] (kernel_init+0x74/0x118) [2.577667] [c0008374] (kernel_init+0x74/0x118) from [c00461b4] (kernel_thread_exit+0x0/0x8) Any idea what went wrong. I am sure about the ioremap() start address, thats what i reserved in ATAGS. Thanks Sandeep On Thu, Feb 28, 2013 at 10:30 AM, sandeep kumar coolsandyfor...@gmail.com wrote: 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. Sounds good, i am gonna try this and let you know.. :) On Wed, Feb 27, 2013 at 8:19 PM, buyitian bu...@live.cn wrote: From: bu...@live.cn To: coolsandyfor...@gmail.com; kernelnewbies@kernelnewbies.org Subject: RE: How to measure the RAM read/write performance Date: Wed, 27 Feb 2013 22:33:15 +0800 CC: dhyla...@gmail.com From: coolsandyfor...@gmail.com Date: Tue, 26 Feb 2013 17:01:54 +0530 Subject: How to measure the RAM read/write performance To: kernelnewbies@kernelnewbies.org CC: dhyla...@gmail.com Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? try to reserve a large region from bootloader(L4 in Qualcomm platform), let's say it is 10MB continuous physical memory. sorry, to be accurate, reserve physical memory is done by kernel cmdline, this cmdline parameter can be passed from L4 to kernel, or configed by kernel itself. the cmdline will be like below: mem=510M@0 test_region=2M@510M above example tells kernel you have totally 512MB physical memory, but kernel will only use the first 510MB, the latter 2MB memory is used by you. how to map and use this region depends on you. in kernel, map this region to an continuous virtual region, note that the pgprot should be uncachable since you want to test without cache. 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail
Re: How to measure the RAM read/write performance
/0x8) [2.449462] Code: e58d4000 e3033fff ea07 e59d2000 (e7952002) [2.455566] ---[ end trace f76f3c76dcb9b9ef ]--- [2.460144] Kernel panic - not syncing: Attempted to kill init! [2.466064] [c004ad10] (unwind_backtrace+0x0/0x12c) from [c064af70] (panic+0x90/0x1bc) [2.474304] [c064af70] (panic+0x90/0x1bc) from [c00eb8dc] (do_exit+0xb8/0x734) [2.481842] [c00eb8dc] (do_exit+0xb8/0x734) from [c0048f8c] (die+0x208/0x23c) [2.489318] [c0048f8c] (die+0x208/0x23c) from [c004e158] (__do_kernel_fault+0x64/0x84) [2.497558] [c004e158] (__do_kernel_fault+0x64/0x84) from [c004e3e0] (do_page_fault+0x268/0x288) [2.506683] [c004e3e0] (do_page_fault+0x268/0x288) from [c003f270] (do_DataAbort+0x34/0x94) [2.515350] [c003f270] (do_DataAbort+0x34/0x94) from [c0044dac] (__dabt_svc+0x4c/0x60) [2.523590] Exception stack(0xe982bf50 to 0xe982bf98) [2.528625] bf40: 0063 c07ea2cf 3ffd 3fff [2.536804] bf60: ea828000 c00461b4 [2.544952] bf80: e982bf98 c0022ed0 c0022ee0 8013 [2.551544] [c0044dac] (__dabt_svc+0x4c/0x60) from [c0022ee0] (sand_misc_init+0x4c/0xac) [2.559967] [c0022ee0] (sand_misc_init+0x4c/0xac) from [c003f3fc] (do_one_initcall+0xd0/0x1a4) [2.568908] [c003f3fc] (do_one_initcall+0xd0/0x1a4) from [c0008374] (kernel_init+0x74/0x118) [2.577667] [c0008374] (kernel_init+0x74/0x118) from [c00461b4] (kernel_thread_exit+0x0/0x8) Any idea what went wrong. I am sure about the ioremap() start address, thats what i reserved in ATAGS. Thanks Sandeep On Thu, Feb 28, 2013 at 10:30 AM, sandeep kumar coolsandyfor...@gmail.com wrote: 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. Sounds good, i am gonna try this and let you know.. :) On Wed, Feb 27, 2013 at 8:19 PM, buyitian bu...@live.cn wrote: From: bu...@live.cn To: coolsandyfor...@gmail.com; kernelnewbies@kernelnewbies.org Subject: RE: How to measure the RAM read/write performance Date: Wed, 27 Feb 2013 22:33:15 +0800 CC: dhyla...@gmail.com From: coolsandyfor...@gmail.com Date: Tue, 26 Feb 2013 17:01:54 +0530 Subject: How to measure the RAM read/write performance To: kernelnewbies@kernelnewbies.org CC: dhyla...@gmail.com Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? try to reserve a large region from bootloader(L4 in Qualcomm platform), let's say it is 10MB continuous physical memory. sorry, to be accurate, reserve physical memory is done by kernel cmdline, this cmdline parameter can be passed from L4 to kernel, or configed by kernel itself. the cmdline will be like below: mem=510M@0 test_region=2M@510M above example tells kernel you have totally 512MB physical memory, but kernel will only use the first 510MB, the latter 2MB memory is used by you. how to map and use this region depends on you. in kernel, map this region to an continuous virtual region, note that the pgprot should be uncachable since you want to test without cache. 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. once you configed like this, you can read/write to this vitual region without data cache invovled. -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- With regards, Sandeep Kumar Anantapalli, -- With regards, Sandeep Kumar Anantapalli, -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list
Re: How to measure the RAM read/write performance
/0x84) from [c004e3e0] (do_page_fault+0x268/0x288) [2.506683] [c004e3e0] (do_page_fault+0x268/0x288) from [c003f270] (do_DataAbort+0x34/0x94) [2.515350] [c003f270] (do_DataAbort+0x34/0x94) from [c0044dac] (__dabt_svc+0x4c/0x60) [2.523590] Exception stack(0xe982bf50 to 0xe982bf98) [2.528625] bf40: 0063 c07ea2cf 3ffd 3fff [2.536804] bf60: ea828000 c00461b4 [2.544952] bf80: e982bf98 c0022ed0 c0022ee0 8013 [2.551544] [c0044dac] (__dabt_svc+0x4c/0x60) from [c0022ee0] (sand_misc_init+0x4c/0xac) [2.559967] [c0022ee0] (sand_misc_init+0x4c/0xac) from [c003f3fc] (do_one_initcall+0xd0/0x1a4) [2.568908] [c003f3fc] (do_one_initcall+0xd0/0x1a4) from [c0008374] (kernel_init+0x74/0x118) [2.577667] [c0008374] (kernel_init+0x74/0x118) from [c00461b4] (kernel_thread_exit+0x0/0x8) Any idea what went wrong. I am sure about the ioremap() start address, thats what i reserved in ATAGS. Thanks Sandeep On Thu, Feb 28, 2013 at 10:30 AM, sandeep kumar coolsandyfor...@gmail.com wrote: 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. Sounds good, i am gonna try this and let you know.. :) On Wed, Feb 27, 2013 at 8:19 PM, buyitian bu...@live.cn wrote: From: bu...@live.cn To: coolsandyfor...@gmail.com; kernelnewbies@kernelnewbies.org Subject: RE: How to measure the RAM read/write performance Date: Wed, 27 Feb 2013 22:33:15 +0800 CC: dhyla...@gmail.com From: coolsandyfor...@gmail.com Date: Tue, 26 Feb 2013 17:01:54 +0530 Subject: How to measure the RAM read/write performance To: kernelnewbies@kernelnewbies.org CC: dhyla...@gmail.com Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? try to reserve a large region from bootloader(L4 in Qualcomm platform), let's say it is 10MB continuous physical memory. sorry, to be accurate, reserve physical memory is done by kernel cmdline, this cmdline parameter can be passed from L4 to kernel, or configed by kernel itself. the cmdline will be like below: mem=510M@0 test_region=2M@510M above example tells kernel you have totally 512MB physical memory, but kernel will only use the first 510MB, the latter 2MB memory is used by you. how to map and use this region depends on you. in kernel, map this region to an continuous virtual region, note that the pgprot should be uncachable since you want to test without cache. 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. once you configed like this, you can read/write to this vitual region without data cache invovled. -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- With regards, Sandeep Kumar Anantapalli, -- With regards, Sandeep Kumar Anantapalli, -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
On Tue, Feb 26, 2013 at 6:50 PM, ankur dwivedi ankurengg2...@gmail.com wrote: I am not sure but what if you make the variable as a volatile? Volatile in simple terms == Consider this eg: 1 int main(void){ 2 int a, x, y; 3 x=a; 4 y=a; 5 } Line 3 with be converted to a LDR instruction. Line 4 is where the compiler optimization comes in. It may be loaded from the previous loaded register. Line 3: ldr r0[r1]; r1 is having the address of a Line 4: mov r2 r0; because we have value of a already in r0. But if a is volatile variable, then compiler places ldr for both instructions(ie for x=a and y = a). so line 4 will change as follows. Line 3: ldr r0[r1];r1 is having the address of a Line 4: ldr r2[r1]; r1 is having the address of a But if cache is enabled, ldr can fetch the data from cache if present. So volatile cannot help us here. Thanks, Arun On Tue, Feb 26, 2013 at 5:01 PM, sandeep kumar coolsandyfor...@gmail.com wrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- Thanks Ankur Dwivedi ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
I am not sure but what if you make the variable as a volatile? Hey i tried. But still is showing the same time for int volatile int. Thanks Sandeep On Tue, Feb 26, 2013 at 6:50 PM, ankur dwivedi ankurengg2...@gmail.comwrote: I am not sure but what if you make the variable as a volatile? On Tue, Feb 26, 2013 at 5:01 PM, sandeep kumar coolsandyfor...@gmail.comwrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- Thanks Ankur Dwivedi -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
let' see what if you do read and write pattern, in certain order so that it will be invalidated by the L1/L2/L3 cache everytime? And how do you suggest we do that?? That is infact my question.. On Tue, Feb 26, 2013 at 9:05 PM, Mulyadi Santosa mulyadi.sant...@gmail.comwrote: On Tue, Feb 26, 2013 at 6:31 PM, sandeep kumar coolsandyfor...@gmail.com wrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. let' see what if you do read and write pattern, in certain order so that it will be invalidated by the L1/L2/L3 cache everytime? AFAIK, one thing for sure, reading data from sequentially and re-read them will make end up reading cache in the 2nd operation and so on. I think the most certain way to do it is to read data (or write) data bigger than total L1/L2/L3 cache. -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
volatile only say to the compiler to not put the variable in a register. Could i ask why does you want to mesure the speed of the RAM ? Is it to study the NUMA behavior of the machine? In development phase of the board, we are trying to measure RAM performance gain while changing type of the RAM. The standard benchmark tools are giving us the Cache performance only. So we want to try some method to measure RAM performance. On Wed, Feb 27, 2013 at 2:30 PM, Karaoui mohamed lamine mohar...@gmail.comwrote: 2013/2/27 sandeep kumar coolsandyfor...@gmail.com I am not sure but what if you make the variable as a volatile? Hey i tried. But still is showing the same time for int volatile int. volatile only say to the compiler to not put the variable in a register. Could i ask why does you want to mesure the speed of the RAM ? Is it to study the NUMA behavior of the machine? Thanks Sandeep On Tue, Feb 26, 2013 at 6:50 PM, ankur dwivedi ankurengg2...@gmail.comwrote: I am not sure but what if you make the variable as a volatile? On Tue, Feb 26, 2013 at 5:01 PM, sandeep kumar coolsandyfor...@gmail.com wrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- Thanks Ankur Dwivedi -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
On Tue, Feb 26, 2013 at 5:01 PM, sandeep kumar coolsandyfor...@gmail.com wrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. Try below points: 1. I am not sure about snapdragon(is it Qualcomm?) but try CONFIG_CPU_DCACHE_DISABLE. 2. You are better off programming some DMA master to do large (and uncached) reads/writes to RAM and timing that. However simple uncached LDR/STR from the CPU may not be a great measure of RAM controller perf. You should always add arm mailing list and please mention the chip set details. cat /proc/cpuinfo is a great way How can i achieve this...Any ideas/suggestions...? -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
On Wed, Feb 27, 2013 at 6:16 PM, sandeep kumar coolsandyfor...@gmail.com wrote: Try below points: 1. I am not sure about snapdragon(is it Qualcomm?) but try CONFIG_CPU_DCACHE_DISABLE. I did that already, device is not booting. In our architecture cache is tightly coupled with CPU. Should be discussed with Qualcomm guys here as to the reason why it doesn't boot up. 2. You are better off programming some DMA master to do large (and uncached) reads/writes to RAM and timing that. DMA is not a standard way, i suppose, as it depends on what I/O peripheral we are doing the DMA. You are doing I/O to RAM so it should do what you want. CPU-DMA CONTROLLER-RAM You should always add arm mailing list and please mention the chip set details. cat /proc/cpuinfo is a great way Processor : ARMv7 Processor rev 2 (v7l) BogoMIPS: 163.38 Features: swp half thumb fastmult vfp edsp neon vfpv3 tls CPU implementer : 0x51 CPU architecture: 7 CPU variant : 0x1 CPU part: 0x00f CPU revision: 2 Thanks Sandeep On Wed, Feb 27, 2013 at 5:37 PM, anish singh anish198519851...@gmail.com wrote: On Tue, Feb 26, 2013 at 5:01 PM, sandeep kumar coolsandyfor...@gmail.com wrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. Try below points: 1. I am not sure about snapdragon(is it Qualcomm?) but try CONFIG_CPU_DCACHE_DISABLE. 2. You are better off programming some DMA master to do large (and uncached) reads/writes to RAM and timing that. However simple uncached LDR/STR from the CPU may not be a great measure of RAM controller perf. You should always add arm mailing list and please mention the chip set details. cat /proc/cpuinfo is a great way How can i achieve this...Any ideas/suggestions...? -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
RE: How to measure the RAM read/write performance
From: coolsandyfor...@gmail.com Date: Tue, 26 Feb 2013 17:01:54 +0530 Subject: How to measure the RAM read/write performance To: kernelnewbies@kernelnewbies.org CC: dhyla...@gmail.com Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? try to reserve a large region from bootloader(L4 in Qualcomm platform), let's say it is 10MB continuous physical memory. in kernel, map this region to an continuous virtual region, note that the pgprot should be uncachable since you want to test without cache. once you configed like this, you can read/write to this vitual region without data cache invovled. -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
RE: How to measure the RAM read/write performance
From: bu...@live.cn To: coolsandyfor...@gmail.com; kernelnewbies@kernelnewbies.org Subject: RE: How to measure the RAM read/write performance Date: Wed, 27 Feb 2013 22:33:15 +0800 CC: dhyla...@gmail.com From: coolsandyfor...@gmail.com Date: Tue, 26 Feb 2013 17:01:54 +0530 Subject: How to measure the RAM read/write performance To: kernelnewbies@kernelnewbies.org CC: dhyla...@gmail.com Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? try to reserve a large region from bootloader(L4 in Qualcomm platform), let's say it is 10MB continuous physical memory. sorry, to be accurate, reserve physical memory is done by kernel cmdline, this cmdline parameter can be passed from L4 to kernel, or configed by kernel itself. the cmdline will be like below: mem=510M@0 test_region=2M@510M above example tells kernel you have totally 512MB physical memory, but kernel will only use the first 510MB, the latter 2MB memory is used by you. how to map and use this region depends on you. in kernel, map this region to an continuous virtual region, note that the pgprot should be uncachable since you want to test without cache. 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. once you configed like this, you can read/write to this vitual region without data cache invovled. -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
On Wed, 27 Feb 2013 15:38:00 +0530, sandeep kumar said: In development phase of the board, we are trying to measure RAM performance gain while changing type of the RAM. The standard benchmark tools are giving us the Cache performance only. So we want to try some method to measure RAM performance. The fact that you can't measure the effect of RAM speed because the L1/2/3 cache masks the effect should tell you something :) If you are seeing a 98% hit rate or so, RAM speed will indeed not matter much. If you're seeing a poor cache hit ratio, you're most likely to get better performance not by changing the RAM, but changing the application to improve its cache usage. And of course, if the application's design is one that is resistant to improved cache hit ratios, it is important you measure RAM performance *with that application running*, not a benchmark. This is because if your application is managing to thrash the cache, the resulting RAM access patterns will be *highly* sensitive to actual program behavior, and any corner cases in the hardware may or may not be hit by the benchmark the same way the application does. pgpZ2odHm7mss.pgp Description: PGP signature ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. Sounds good, i am gonna try this and let you know.. :) On Wed, Feb 27, 2013 at 8:19 PM, buyitian bu...@live.cn wrote: From: bu...@live.cn To: coolsandyfor...@gmail.com; kernelnewbies@kernelnewbies.org Subject: RE: How to measure the RAM read/write performance Date: Wed, 27 Feb 2013 22:33:15 +0800 CC: dhyla...@gmail.com From: coolsandyfor...@gmail.com Date: Tue, 26 Feb 2013 17:01:54 +0530 Subject: How to measure the RAM read/write performance To: kernelnewbies@kernelnewbies.org CC: dhyla...@gmail.com Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? try to reserve a large region from bootloader(L4 in Qualcomm platform), let's say it is 10MB continuous physical memory. sorry, to be accurate, reserve physical memory is done by kernel cmdline, this cmdline parameter can be passed from L4 to kernel, or configed by kernel itself. the cmdline will be like below: mem=510M@0 test_region=2M@510M above example tells kernel you have totally 512MB physical memory, but kernel will only use the first 510MB, the latter 2MB memory is used by you. how to map and use this region depends on you. in kernel, map this region to an continuous virtual region, note that the pgprot should be uncachable since you want to test without cache. 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. once you configed like this, you can read/write to this vitual region without data cache invovled. -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
6c632065 6e69736f 20612067 62206d76 696b6361 6120676e 6c61206e [2.241119] [2.241119] R5: 0xea827f80: [2.245361] 7f80 [2.253509] 7fa0 [2.261688] 7fc0 [2.269836] 7fe0 [2.278015] 8000 [2.286163] 8020 [2.294311] 8040 [2.302490] 8060 [2.310638] [2.310638] R6: 0xc0046134: [2.314880] 6134 eb038dff eb02a1f9 e1a03007 e1a5 e1a02006 eb025566 e59f3008 e5834000 [2.323059] 6154 e8bd41f0 ea0259b3 c09059d4 c08fbcd4 c064e808 c0791c0d c0958220 e59fc020 [2.331207] 6174 e92d4007 e59f301c e59cc000 e1a02001 e59f1014 e58dc000 eb055aab e3a0 [2.339385] 6194 e8bd800e c0958220 c0791c27 c0791c1b e121f007 e1a4 e1a0e006 e1a0f005 [2.347534] 61b4 eb02959a e320f000 e59f300c e5932000 e2822001 e5832000 e12fff1e c0958224 [2.355682] 61d4 e59f300c e5932000 e2422001 e5832000 e12fff1e c0958224 e12fff1e e12fff1e [2.363861] 61f4 e12fff1e e12fff1e eafe e59f3014 e92d4010 e5933004 e353 08bd8010 [2.372009] 6214 e12fff33 e8bd8010 c0958224 e59f3014 e1a01000 e92d4010 e5d3 e1a0e00f [2.380187] Process swapper (pid: 1, stack limit = 0xe982a2e8) [2.385986] Stack: (0xe982bf98 to 0xe982c000) [2.390319] bf80: 3ffd a27e [2.398498] bfa0: c0037b4c c0022e94 c003f3fc e9814a80 00373231 [2.406646] bfc0: 0236 c0037b4c c003818c c00461b4 0013 [2.414825] bfe0: c0008374 c0008300 c00461b4 c00461b4 [2.423004] [c0022ee0] (sand_misc_init+0x4c/0xac) from [c003f3fc] (do_one_initcall+0xd0/0x1a4) [2.431915] [c003f3fc] (do_one_initcall+0xd0/0x1a4) from [c0008374] (kernel_init+0x74/0x118) [2.440673] [c0008374] (kernel_init+0x74/0x118) from [c00461b4] (kernel_thread_exit+0x0/0x8) [2.449462] Code: e58d4000 e3033fff ea07 e59d2000 (e7952002) [2.455566] ---[ end trace f76f3c76dcb9b9ef ]--- [2.460144] Kernel panic - not syncing: Attempted to kill init! [2.466064] [c004ad10] (unwind_backtrace+0x0/0x12c) from [c064af70] (panic+0x90/0x1bc) [2.474304] [c064af70] (panic+0x90/0x1bc) from [c00eb8dc] (do_exit+0xb8/0x734) [2.481842] [c00eb8dc] (do_exit+0xb8/0x734) from [c0048f8c] (die+0x208/0x23c) [2.489318] [c0048f8c] (die+0x208/0x23c) from [c004e158] (__do_kernel_fault+0x64/0x84) [2.497558] [c004e158] (__do_kernel_fault+0x64/0x84) from [c004e3e0] (do_page_fault+0x268/0x288) [2.506683] [c004e3e0] (do_page_fault+0x268/0x288) from [c003f270] (do_DataAbort+0x34/0x94) [2.515350] [c003f270] (do_DataAbort+0x34/0x94) from [c0044dac] (__dabt_svc+0x4c/0x60) [2.523590] Exception stack(0xe982bf50 to 0xe982bf98) [2.528625] bf40: 0063 c07ea2cf 3ffd 3fff [2.536804] bf60: ea828000 c00461b4 [2.544952] bf80: e982bf98 c0022ed0 c0022ee0 8013 [2.551544] [c0044dac] (__dabt_svc+0x4c/0x60) from [c0022ee0] (sand_misc_init+0x4c/0xac) [2.559967] [c0022ee0] (sand_misc_init+0x4c/0xac) from [c003f3fc] (do_one_initcall+0xd0/0x1a4) [2.568908] [c003f3fc] (do_one_initcall+0xd0/0x1a4) from [c0008374] (kernel_init+0x74/0x118) [2.577667] [c0008374] (kernel_init+0x74/0x118) from [c00461b4] (kernel_thread_exit+0x0/0x8) Any idea what went wrong. I am sure about the ioremap() start address, thats what i reserved in ATAGS. Thanks Sandeep On Thu, Feb 28, 2013 at 10:30 AM, sandeep kumar coolsandyfor...@gmail.comwrote: 1. use early_param to get the physical start address and size of test_region, or you can just ignore this step and hard code to 510M and 2M for test purpose only. 2. use ioremap_nocache() to map this region to a virtual region. note that this funtion may fail if you are asking a very large vitual memroy region. Sounds good, i am gonna try this and let you know.. :) On Wed, Feb 27, 2013 at 8:19 PM, buyitian bu...@live.cn wrote: From: bu...@live.cn To: coolsandyfor...@gmail.com; kernelnewbies@kernelnewbies.org Subject: RE: How to measure the RAM read/write performance Date: Wed, 27 Feb 2013 22:33:15 +0800 CC: dhyla...@gmail.com From: coolsandyfor...@gmail.com Date: Tue, 26 Feb 2013 17:01:54 +0530 Subject: How to measure the RAM read/write performance To: kernelnewbies@kernelnewbies.org CC
How to measure the RAM read/write performance
Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
The mechanism to disable the L1/L2 memory caches of the CPU will depend on the processor architecture itself. This can be possibly be done at kernel boot-time or more preferably, in your case, through a kernel module. For Intel x86, the kernel module would look something like following: (The key is setting bit 30 of the control register cr0) mov eax,cr0 or eax, 4000H ;set bit 30 without clearing the others mov cr0,eax Reference: http://stackoverflow.com/questions/1108485/disable-l2-l1-caches http://dirac.org/linux/writing/lkmpg/2.6/lkmpg-2.6.0.html ~ Gaurav On Tue, Feb 26, 2013 at 12:31 PM, sandeep kumar coolsandyfor...@gmail.comwrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- Gaurav Jain Associate Software Engineer VxVM Escalations Team, SAMG Symantec Software India Pvt. Ltd. ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
I am not sure but what if you make the variable as a volatile? On Tue, Feb 26, 2013 at 5:01 PM, sandeep kumar coolsandyfor...@gmail.comwrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. How can i achieve this...Any ideas/suggestions...? -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies -- Thanks Ankur Dwivedi ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
On Tue, Feb 26, 2013 at 6:31 PM, sandeep kumar coolsandyfor...@gmail.com wrote: Hi All In performance benchmark tools, When we profile read/write timings mostly, those read/writes are done to cache only. I want to measure my DDR(RAM chip) performance. So i want to make sure, every read/write should happen to DDR RAM chip only. let' see what if you do read and write pattern, in certain order so that it will be invalidated by the L1/L2/L3 cache everytime? AFAIK, one thing for sure, reading data from sequentially and re-read them will make end up reading cache in the 2nd operation and so on. I think the most certain way to do it is to read data (or write) data bigger than total L1/L2/L3 cache. -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
On Tue, 26 Feb 2013 22:35:35 +0700, Mulyadi Santosa said: let' see what if you do read and write pattern, in certain order so that it will be invalidated by the L1/L2/L3 cache everytime? AFAIK, one thing for sure, reading data from sequentially and re-read them will make end up reading cache in the 2nd operation and so on. I think the most certain way to do it is to read data (or write) data bigger than total L1/L2/L3 cache. Of you could just download a copy of memtest+ and run that - I think that provides some timing info in addition to actually testing your memory. pgpXhrIunMesq.pgp Description: PGP signature ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
On Tue, Feb 26, 2013 at 11:57 PM, valdis.kletni...@vt.edu wrote: Of you could just download a copy of memtest+ and run that - I think that provides some timing info in addition to actually testing your memory. How could I missed memtest ! :) Of course, that's the easiest way I could think of... if the OP doesn't mind with non programming method IMHO Great idea! -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
Re: How to measure the RAM read/write performance
How could I missed memtest ! :) Of course, that's the easiest way I could think of... if the OP doesn't mind with non programming method IMHO But i want to check it on ARM based Android board.. On Wed, Feb 27, 2013 at 6:12 AM, Mulyadi Santosa mulyadi.sant...@gmail.comwrote: On Tue, Feb 26, 2013 at 11:57 PM, valdis.kletni...@vt.edu wrote: Of you could just download a copy of memtest+ and run that - I think that provides some timing info in addition to actually testing your memory. How could I missed memtest ! :) Of course, that's the easiest way I could think of... if the OP doesn't mind with non programming method IMHO Great idea! -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com -- With regards, Sandeep Kumar Anantapalli, ___ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies