Re: [Xen-devel] [PATCH 4/4] tools/xenalyze: Allow automatic resizing of sample buffers

[George Dunlap <george.dunlap@xxxxxxxxxx>]

On Mon, Aug 8, 2016 at 6:11 PM, anshul makkar <anshul.makkar@xxxxxxxxxx> wrote:
> On 08/08/16 10:54, George Dunlap wrote:
>>
>> Rather than have large fixed-size buffers, start with smaller buffers
>> and allow them to grow as needed (doubling each time), with a fairly
>> large maximum.  Allow this maximum to be set by a command-line
>> parameter.
>>
>> Signed-off-by: George Dunlap <george.dunlap@xxxxxxxxxx>
>> ---
>> CC: Ian Jackson <ian.jackson@xxxxxxxxxx>
>> CC: Wei Liu <wei.liu2@xxxxxxxxxx>
>> CC: Dario Faggioli <dario.faggioli@xxxxxxxxxx>
>> CC: Anshul Makkar <anshul.makkar@xxxxxxxxxx>
>> ---
>>   tools/xentrace/xenalyze.c | 95
>> +++++++++++++++++++++++++++++++++--------------
>>   1 file changed, 68 insertions(+), 27 deletions(-)
>>
>> diff --git a/tools/xentrace/xenalyze.c b/tools/xentrace/xenalyze.c
>> index 455cbdf..a4d8823 100644
>> --- a/tools/xentrace/xenalyze.c
>> +++ b/tools/xentrace/xenalyze.c
>> @@ -44,7 +44,8 @@ struct mread_ctrl;
>>   #define QHZ_FROM_HZ(_hz) (((_hz) << 10)/ 1000000000)
>>
>>   #define ADDR_SPACE_BITS 48
>> -#define DEFAULT_SAMPLE_SIZE 10240
>> +#define DEFAULT_SAMPLE_SIZE 1024
>> +#define DEFAULT_SAMPLE_MAX  1024*1024*32
>>   #define DEFAULT_INTERVAL_LENGTH 1000
>>
>
>>       s->event_count++;
>>
>>       if (!c)
>>           return;
>>
>>       if(opt.sample_size) {
>> -        int lap = (s->count/opt.sample_size)+1,
>> -            index =s->count % opt.sample_size;
>> -        if((index - (lap/3))%lap == 0) {
>> -            if(!s->sample) {
>> -                s->sample = malloc(sizeof(*s->sample) * opt.sample_size);
>> -                if(!s->sample) {
>> -                    fprintf(stderr, "%s: malloc failed!\n", __func__);
>> -                    error(ERR_SYSTEM, NULL);
>> -                }
>> +        if (s->count >= s->sample_size
>> +            && (s->count == 0
>> +                || opt.sample_max == 0
>> +                || s->sample_size < opt.sample_max)) {
>> +            int new_size;
>> +            void * new_sample = NULL;
>> +
>> +            new_size = s->sample_size << 1;
>
> Sorry for my ignorance here, but why we have chosen to double the size.
> Can't we increase by fixed size X where X < double size. Are we sure that we
> will be able to fully utilize the double sized buffer.

Well we're sure it will be at least 50% utilized. :-)  Also remember
that we're in userspace here, so no physical memory will be allocated
until the process actually writes to the virtual memory anyway.  The
main goal is to balance virtual memory usage against the cost of
reallocation (which will probably involve copying all the data from
the old buffer into the new buffer), and there's no point in having a
really complicated algorithm until we find cases where the simple
algorithm is a problem.

 -George

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel