Mike Xu Mike Xu - 5 months ago 19
Linux Question

Weird program latency behavior on VM

I wrote a program to read 256KB array to get 1ms latency. The program is pretty simple and attached.
However, when I run it on VM on Xen, I found that the latency is not stable. It has the following pattern: The time unit is ms.

#totalCycle CyclePerLine totalms
22583885 5513 6.452539
3474342 848 0.992669
3208486 783 0.916710
25848572 6310 7.385306
3225768 787 0.921648
3210487 783 0.917282
25974700 6341 7.421343
3244891 792 0.927112
3276027 799 0.936008
25641513 6260 7.326147
3531084 862 1.008881
3233687 789 0.923911
22397733 5468 6.399352
3523403 860 1.006687
3586178 875 1.024622
26094384 6370 7.455538
3540329 864 1.011523
3812086 930 1.089167
25907966 6325 7.402276


I'm thinking some process is doing something and it's like an event-driven process. Does any one encounter this before? or anyone can point out the potential process/services that could make this happen?

Below is my program. I run it for 1000 times. Each time got the one line of the result above.

#include <iostream>
#include <cstdio>
#include <cstdlib>
#include <string>
#include <ctime>

using namespace std;

#if defined(__i386__)
static __inline__ unsigned long long rdtsc(void)
{
unsigned long long int x;
__asm__ volatile (".byte 0x0f, 0x31" : "=A" (x));
return x;
}
#elif defined(__x86_64__)
static __inline__ unsigned long long rdtsc(void)
{
unsigned hi, lo;
__asm__ __volatile__ ("rdtsc" : "=a"(lo), "=d"(hi));
return ( (unsigned long long)lo)|( ((unsigned long long)hi)<<32 );
}
#endif

#define CACHE_LINE_SIZE 64

#define WSS 24567 /* 24 Mb */
#define NUM_VARS WSS * 1024 / sizeof(long)

#define KHZ 3500000

// ./a.out memsize(in KB)
int main(int argc, char** argv)
{
unsigned long wcet = atol(argv[1]);
unsigned long mem_size_KB = 256; // mem size in KB
unsigned long mem_size_B = mem_size_KB * 1024; // mem size in Byte
unsigned long count = mem_size_B / sizeof(long);
unsigned long row = mem_size_B / CACHE_LINE_SIZE;
int col = CACHE_LINE_SIZE / sizeof(long);

unsigned long long start, finish, dur1;
unsigned long temp;

long *buffer;
buffer = new long[count];

// init array
for (unsigned long i = 0; i < count; ++i)
buffer[i] = i;

for (unsigned long i = row-1; i >0; --i) {
temp = rand()%i;
swap(buffer[i*col], buffer[temp*col]);
}

// warm the cache again
temp = buffer[0];
for (unsigned long i = 0; i < row-1; ++i) {
temp = buffer[temp];
}

// First read, should be cache hit
temp = buffer[0];
start = rdtsc();
int sum = 0;
for(int wcet_i = 0; wcet_i < wcet; wcet_i++)
{
for(int j=0; j<21; j++)
{
for (unsigned long i = 0; i < row-1; ++i) {
if (i%2 == 0) sum += buffer[temp];
else sum -= buffer[temp];
temp = buffer[temp];
}
}
}
finish = rdtsc();
dur1 = finish-start;

// Res
printf("%lld %lld %.6f\n", dur1, dur1/row, dur1*1.0/KHZ);
delete[] buffer;
return 0;
}

Answer

The use of the RDTSC instruction in a virtual machine is complicated. It is likely that the hypervisor (Xen) is emulating the RDTSC instruction by trapping it. Your fastest runs show around 800 cycles/cache line, which is very, very, slow... the only explanation is that the RDTSC results in a trap that is handled by the hypervisor, that overhead is a performance bottleneck. I'm not sure about the even longer time that you see periodically, but given that the RDTSC is being trapped, all timing bets are off.

You can read more about it here

http://xenbits.xen.org/docs/4.2-testing/misc/tscmode.txt

Instructions in the rdtsc family are non-privileged, but privileged software may set a cpuid bit to cause all rdtsc family instructions to trap. This trap can be detected by Xen, which can then transparently "emulate" the results of the rdtsc instruction and return control to the code following the rdtsc instruction

By the way, that article is wrong in that the hypervisor doesn't set a cpuid bit to cause RDTSC to trap, it is bit #2 in Control Register 4 (CR4.TSD):

http://en.wikipedia.org/wiki/Control_register#CR4