Hi Folks, Is there any benchmarking tool to test read/write speed on Linux and which works on ARM ? Regards, --Sagar
On Tue, 30 Sep 2014 02:53:05 +0530, Ssagarr Patil said:
Is there any benchmarking tool to test read/write speed on Linux and which works on ARM ?
This is always a hairy problem. Do you care about the memory speed of the DIMM, or the actual *effective* memory speed? They can be very different, due to the effects of L1/2/3 cache (both positive if there's a cache hit, and negative if you need a cache line load, or if you need a line load that requires a cache line writeback first), whether or not you have a TLB active, and whether *that* has a hit or miss (which varies considerably if you're walking an array and hitting a miss every 4K bytes, versus walking a nest of pointers where you get a TLB miss almost every reference), and a bunch of other issues I lack the caffeine to enumerate... And of course, the *effective* speed (which is what you should probably care about rather than the *rated* speed) will differ based on workload due to the factors listed above....
I use the memtest86+ numbers about memory speed. It gives numbers for RAM and different cache levels. On Mon, Sep 29, 2014 at 11:42 PM, <Valdis.Kletnieks@vt.edu> wrote:
On Tue, 30 Sep 2014 02:53:05 +0530, Ssagarr Patil said:
Is there any benchmarking tool to test read/write speed on Linux and which works on ARM ?
This is always a hairy problem. Do you care about the memory speed of the DIMM, or the actual *effective* memory speed? They can be very different, due to the effects of L1/2/3 cache (both positive if there's a cache hit, and negative if you need a cache line load, or if you need a line load that requires a cache line writeback first), whether or not you have a TLB active, and whether *that* has a hit or miss (which varies considerably if you're walking an array and hitting a miss every 4K bytes, versus walking a nest of pointers where you get a TLB miss almost every reference), and a bunch of other issues I lack the caffeine to enumerate...
And of course, the *effective* speed (which is what you should probably care about rather than the *rated* speed) will differ based on workload due to the factors listed above....
_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
-- Peter
Hi, ----------------------------------------
Date: Tue, 30 Sep 2014 17:49:42 +0200 Subject: Re: Test RAM read/write speed From: peter.senna@gmail.com To: Valdis.Kletnieks@vt.edu CC: hugarsagar@outlook.com; kernelnewbies@kernelnewbies.org
I use the memtest86+ numbers about memory speed. It gives numbers for RAM and different cache levels.
you mean numbers for read/write ?
On Mon, Sep 29, 2014 at 11:42 PM, <Valdis.Kletnieks@vt.edu> wrote:
On Tue, 30 Sep 2014 02:53:05 +0530, Ssagarr Patil said:
Is there any benchmarking tool to test read/write speed on Linux and which works on ARM ?
This is always a hairy problem. Do you care about the memory speed of the DIMM, or the actual *effective* memory speed? They can be very different, due to the effects of L1/2/3 cache (both positive if there's a cache hit, and negative if you need a cache line load, or if you need a line load that requires a cache line writeback first), whether or not you have a TLB active, and whether *that* has a hit or miss (which varies considerably if you're walking an array and hitting a miss every 4K bytes, versus walking a nest of pointers where you get a TLB miss almost every reference), and a bunch of other issues I lack the caffeine to enumerate...
And of course, the *effective* speed (which is what you should probably care about rather than the *rated* speed) will differ based on workload due to the factors listed above....
May be a tool is needed at uboot level to test the speeds.
Thanks. --Sagar
_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
-- Peter
_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
On Wed, 01 Oct 2014 01:32:19 +0530, Ssagarr Patil said:
May be a tool is needed at uboot level to test the speeds.
What will that actually *tell* you though? Or more to the point, what *actionable* information will that give you? The cache speeds and memory speeds are something you're stuck with unless you upgrade the computer. And especially in the ARM SoC world, "upgrade" means "buy the next Raspberry Pi model" or "time for a new Beaglebone" or whatever. Now, what would be *useful* is a tool that measures how much you're thrashing the L1/L2 cache and the TLB - often rewriting the code will allow much better performance by minimizing cache/TLB misses, but you need to know which code is sucking before you can make it un-sucking. Also, that's a *different* question than what the memory speed is - at that point, it's more about how effectively you're using the memory bandwidth.
On Tue, Sep 30, 2014 at 10:02 PM, Ssagarr Patil <hugarsagar@outlook.com> wrote:
Hi,
----------------------------------------
Date: Tue, 30 Sep 2014 17:49:42 +0200 Subject: Re: Test RAM read/write speed From: peter.senna@gmail.com To: Valdis.Kletnieks@vt.edu CC: hugarsagar@outlook.com; kernelnewbies@kernelnewbies.org
I use the memtest86+ numbers about memory speed. It gives numbers for RAM and different cache levels.
you mean numbers for read/write ? I don't know. Check this out: https://plus.google.com/+PeterSenna/posts/iw6p2ZSU8VE
On Mon, Sep 29, 2014 at 11:42 PM, <Valdis.Kletnieks@vt.edu> wrote:
On Tue, 30 Sep 2014 02:53:05 +0530, Ssagarr Patil said:
Is there any benchmarking tool to test read/write speed on Linux and which works on ARM ?
This is always a hairy problem. Do you care about the memory speed of the DIMM, or the actual *effective* memory speed? They can be very different, due to the effects of L1/2/3 cache (both positive if there's a cache hit, and negative if you need a cache line load, or if you need a line load that requires a cache line writeback first), whether or not you have a TLB active, and whether *that* has a hit or miss (which varies considerably if you're walking an array and hitting a miss every 4K bytes, versus walking a nest of pointers where you get a TLB miss almost every reference), and a bunch of other issues I lack the caffeine to enumerate...
And of course, the *effective* speed (which is what you should probably care about rather than the *rated* speed) will differ based on workload due to the factors listed above....
May be a tool is needed at uboot level to test the speeds.
Thanks. --Sagar
_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
-- Peter
_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
-- Peter
participants (3)
-
Peter Senna Tschudin -
Ssagarr Patil -
Valdis.Kletnieks＠vt.edu