Re: [PATCH 0/3] xen: evtchn and gntdev device fixes and perf improvements

[David Vrabel <david.vrabel@citrix.com>]

[-- Attachment #1: Type: text/plain, Size: 778 bytes --]

On 19/07/13 15:51, David Vrabel wrote:
> 
> Patch 3 improves the scalability of the evtchn device when it is used
> by multiple processes (e.g., multiple qemus).  As you can see from the
> graph[1] it is a signficant improvement but still less than ideal.  I
> suspect that this may be due to the per-domain event lock inside Xen
> rather than anything kernel-side.
> 
> The graphed data was collected from dom0 with 8 VCPUs on a host with 8
> CPUs.

Attached is the (slightly cheesy) test program I used to generate the
results.  This also triggered the deadlock fixed by patch 1.

It spawns N threads each of which opens /dev/xen/evtchn channel sets up
an event channel with both ends in the same domain.  The event channels
are manually distributed between the VCPUs.

David

[-- Attachment #2: evtchn-stress.c --]
[-- Type: text/x-csrc, Size: 2630 bytes --]

#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>

#include <xenctrl.h>

#define MAX_THREADS 128
#define MAX_EVENTS 100000

#define FIRST_IRQ 62
#define MAX_VCPUS 8

static unsigned long long elapsed[MAX_THREADS];

static void bind_irq(unsigned port, unsigned cpu)
{
    char buf[64];
    FILE *f;
    unsigned irq = port + FIRST_IRQ;

    snprintf(buf, sizeof(buf), "/proc/irq/%d/smp_affinity", irq);

    f = fopen(buf, "w");
    if (!f)
        return;

    snprintf(buf, sizeof(buf), "%x", 1 << cpu);
    fwrite(buf, strlen(buf), 1, f);
    fclose(f);
}

void *test_thread(void *arg)
{
    unsigned i = (intptr_t)arg;
    xc_evtchn *xce;
    evtchn_port_t lport;
    evtchn_port_t rport;
    unsigned events = 0;
    struct timespec start, end;

    xce = xc_evtchn_open(NULL, 0);
    if (!xce)
        return NULL;

    lport = xc_evtchn_bind_unbound_port(xce, 0);
    if (lport < 0)
        return NULL;

    rport = xc_evtchn_bind_interdomain(xce, 0, lport);

    bind_irq(lport, lport % MAX_VCPUS);
    bind_irq(rport, rport % MAX_VCPUS);

    xc_evtchn_notify(xce, lport);

    clock_gettime(CLOCK_MONOTONIC, &start);

    while (events < MAX_EVENTS) {
        evtchn_port_t port;

        port = xc_evtchn_pending(xce);
        if (port == lport)
            xc_evtchn_notify(xce, lport);
        else
            xc_evtchn_notify(xce, rport);
        xc_evtchn_unmask(xce, port);

        events++;
    }

    clock_gettime(CLOCK_MONOTONIC, &end);

    xc_evtchn_close(xce);

    elapsed[i] = (end.tv_sec - start.tv_sec) * 1000000000ull
        + (end.tv_nsec - start.tv_nsec);

    return NULL;
}

static pthread_t threads[MAX_THREADS];

int start_thread(unsigned i)
{
    return pthread_create(&threads[i], NULL, test_thread, (void *)(intptr_t)i);
}

void run_test(unsigned num_threads)
{
    unsigned i;
    double accum = 0.0;

    for (i = 0; i < num_threads; i++)
        start_thread(i);

    for (i = 0; i < num_threads; i++)
        pthread_join(threads[i], NULL);

    for (i  = 0; i < num_threads; i++)
        accum += MAX_EVENTS / (double)elapsed[i];

    printf("  num_threads: %u\n", num_threads);
    printf("  rate: %.1f event/s\n", accum / num_threads * 1000000000.0);
}

int main(int argc, char *argv[])
{
    unsigned num_threads;
    unsigned num_iters;
    unsigned i;

    if (argc <=1 )
        fprintf(stderr, "Usage: %s <iters> <threads>\n", argv[0]);

    num_iters = atoi(argv[1]);
    num_threads = atoi(argv[2]);

    for (i = 0; i < num_iters; i++) {
        printf("iter: %d\n", i);
        run_test(num_threads);
    }

    return 0;
}

[-- Attachment #3: Type: text/plain, Size: 126 bytes --]

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