[dpdk-dev] cost of reading tsc register

[Pawel Wodkowski]

On 2015-04-20 16:37, Ravi Kumar Iyer wrote:
> Hi,
> We were doing some code optimizations , running DPDK based applications, and chanced upon the rte_rdtsc function [ to read tsc timestamp register value ] consuming cpu cycles of the order of 100clock cycles with a delta of upto 40cycles at times [ 60-140 cycles]
>
> We are actually building up a cpu intensive application which is also very clock cycle sensitive and this is impacting our implementation.
>
> To validate the same using a small/vanilla application we wrote a small code and tested on a single core.
> Has anyone else faced a similar issue or are we doing something really atrocious here.
>
> Below is the pseudo snip of the same:
>
>
...
>      for (i = 0; i < 8 ; i++)
>      {
>          g_tsc_cost[i] = rte_rdtsc();
>        }
...
>
>      uint64_t sc = rte_rdtsc(); /* start count */
>      test_tsc_cost();
>      uint64_t ec = rte_rdtsc(); /* end count */
>

I am no an expert in this topic but I can share you knowledge I got 
during lib jobstats implementation (I think you can find it useful in 
your case with small modification in getting the time).

The rte_rdtsc() (it is wrapper to asm rdtsc instruction) is pretty 
useless in this particular use case. This instruction is pipelined and 
because of this you wont get precise time.

The same is true for rte_rdtsc_precise(). This one is memory barrier 
followed by rte_rdtsc(). I was surprised that compiler in most cases 
remove the memory barrier on '-Os' and '-O3', so final code might not be 
different than rte_rdtsc().

There is no perfect solution for your problem.
Assuming you want measure pure code execution time you need to use the 
... CPUID instruction :D together with RDTSC and RTDSCP. Yes, this not a 
joke. The CPUID is some kind of barrier to the out-of-order execution. 
Writing this in pseudo code:

static inline uint64_t
rte_rdtscp(void)
{
	union {
		uint64_t tsc_64;
		struct {
			uint32_t lo_32;
			uint32_t hi_32;
		};
	} tsc;

#ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
	/* What ever is needed here */
#endif

	asm volatile("rdtscp" :
		     "=a" (tsc.lo_32),
		     "=d" (tsc.hi_32));
	return tsc.tsc_64;
}

uint_64_t
timestamp_start(void) {
	/* Execution barier */
	asm CPUID;
	return rte_rdtsc(); /* without 'p' */
}

uint_64_t
timestamp_get(void) {
	/* Execution barier */
	uint64_t time = rte_rdtscp(); /* without 'p' */
	asm CPUID
	return time;
}

void
do_some_task(void)
{
	g_tsc_cost[i] = timestamp_get();
}

/* warmup cache */
timestamp_start();
timestamp_start();
timestamp_start();

start_time = timestamp_start();
do_some_task();
end_time = timestamp_get()
...

And some words about performance here:
If you want use it many times in code and measure intervals less than 
few thousands of cycles you will kill your application becouse of 
processor stall at CPUID and RDTSCP instruction so use it wisely.
During l2_fwd_jobstats example implementation I tested those cases.
With original rte_rdtscp() app was able to handle about 64B packets with 
2x7.5GiB traffic/per core. When I used CPUID and RDTSCP to get 
"accurate" timestamps I got max 2x4.5GiB. So again: use it wisely.

And one word that is totally my opinion I came up: those CPUs are no 
designed to do very precise time measurements, because there is no easy 
way to implement it without getting significant performance penalty.

>
> Just to compare, On few bare metal implementations of non-intel processors, we are seeing the similar code print values with a delta of 3-4 cycles and thus its becoming a bit difficult to digest as well.  Grateful for any help/guidance here.
>

I think you should also isolate the CPU from scheduling and use IRQ 
affinity to remove any unwanted interference form system.

-- 
Pawel