[Qemu-devel] [PATCH] iothread: enable AioContext polling by default

[Qemu-devel] [PATCH] iothread: enable AioContext polling by default

[Stefan Hajnoczi]

IOThread AioContexts are likely to consist only of event sources like
virtqueue ioeventfds and LinuxAIO completion eventfds that are pollable
from userspace (without system calls).

We recently merged the AioContext polling feature but didn't enable it
by default yet.  I have gone back over the performance data on the
mailing list and picked a default polling value that gave good results.

Let's enable AioContext polling by default so users don't have another
switch they need to set manually.  If performance regressions are found
we can still disable this for the QEMU 2.9 release.

Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Cc: Karl Rister <krister@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 iothread.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/iothread.c b/iothread.c
index 7bedde8..257b01d 100644
--- a/iothread.c
+++ b/iothread.c
@@ -30,6 +30,12 @@ typedef ObjectClass IOThreadClass;
 #define IOTHREAD_CLASS(klass) \
    OBJECT_CLASS_CHECK(IOThreadClass, klass, TYPE_IOTHREAD)
 
+/* Benchmark results from 2016 on NVMe SSD drives show max polling times around
+ * 16-32 microseconds yield IOPS improvements for both iodepth=1 and iodepth=32
+ * workloads.
+ */
+#define IOTHREAD_POLL_MAX_NS_DEFAULT 32768ULL
+
 static __thread IOThread *my_iothread;
 
 AioContext *qemu_get_current_aio_context(void)
@@ -71,6 +77,13 @@ static int iothread_stop(Object *object, void *opaque)
     return 0;
 }
 
+static void iothread_instance_init(Object *obj)
+{
+    IOThread *iothread = IOTHREAD(obj);
+
+    iothread->poll_max_ns = IOTHREAD_POLL_MAX_NS_DEFAULT;
+}
+
 static void iothread_instance_finalize(Object *obj)
 {
     IOThread *iothread = IOTHREAD(obj);
@@ -215,6 +228,7 @@ static const TypeInfo iothread_info = {
     .parent = TYPE_OBJECT,
     .class_init = iothread_class_init,
     .instance_size = sizeof(IOThread),
+    .instance_init = iothread_instance_init,
     .instance_finalize = iothread_instance_finalize,
     .interfaces = (InterfaceInfo[]) {
         {TYPE_USER_CREATABLE},
-- 
2.9.3

Re: [Qemu-devel] [PATCH] iothread: enable AioContext polling by default

[Paolo Bonzini]

On 26/01/2017 18:01, Stefan Hajnoczi wrote:
> IOThread AioContexts are likely to consist only of event sources like
> virtqueue ioeventfds and LinuxAIO completion eventfds that are pollable
> from userspace (without system calls).
> 
> We recently merged the AioContext polling feature but didn't enable it
> by default yet.  I have gone back over the performance data on the
> mailing list and picked a default polling value that gave good results.
> 
> Let's enable AioContext polling by default so users don't have another
> switch they need to set manually.  If performance regressions are found
> we can still disable this for the QEMU 2.9 release.
> 
> Cc: Paolo Bonzini <pbonzini@redhat.com>
> Cc: Christian Borntraeger <borntraeger@de.ibm.com>
> Cc: Karl Rister <krister@redhat.com>
> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> ---
>  iothread.c | 14 ++++++++++++++
>  1 file changed, 14 insertions(+)
> 
> diff --git a/iothread.c b/iothread.c
> index 7bedde8..257b01d 100644
> --- a/iothread.c
> +++ b/iothread.c
> @@ -30,6 +30,12 @@ typedef ObjectClass IOThreadClass;
>  #define IOTHREAD_CLASS(klass) \
>     OBJECT_CLASS_CHECK(IOThreadClass, klass, TYPE_IOTHREAD)
>  
> +/* Benchmark results from 2016 on NVMe SSD drives show max polling times around
> + * 16-32 microseconds yield IOPS improvements for both iodepth=1 and iodepth=32
> + * workloads.
> + */
> +#define IOTHREAD_POLL_MAX_NS_DEFAULT 32768ULL
> +
>  static __thread IOThread *my_iothread;
>  
>  AioContext *qemu_get_current_aio_context(void)
> @@ -71,6 +77,13 @@ static int iothread_stop(Object *object, void *opaque)
>      return 0;
>  }
>  
> +static void iothread_instance_init(Object *obj)
> +{
> +    IOThread *iothread = IOTHREAD(obj);
> +
> +    iothread->poll_max_ns = IOTHREAD_POLL_MAX_NS_DEFAULT;
> +}
> +
>  static void iothread_instance_finalize(Object *obj)
>  {
>      IOThread *iothread = IOTHREAD(obj);
> @@ -215,6 +228,7 @@ static const TypeInfo iothread_info = {
>      .parent = TYPE_OBJECT,
>      .class_init = iothread_class_init,
>      .instance_size = sizeof(IOThread),
> +    .instance_init = iothread_instance_init,
>      .instance_finalize = iothread_instance_finalize,
>      .interfaces = (InterfaceInfo[]) {
>          {TYPE_USER_CREATABLE},
> 

ACK

Paolo

Re: [Qemu-devel] [PATCH] iothread: enable AioContext polling by default

[Stefan Hajnoczi]

[-- Attachment #1: Type: text/plain, Size: 1058 bytes --]

On Thu, Jan 26, 2017 at 05:01:19PM +0000, Stefan Hajnoczi wrote:
> IOThread AioContexts are likely to consist only of event sources like
> virtqueue ioeventfds and LinuxAIO completion eventfds that are pollable
> from userspace (without system calls).
> 
> We recently merged the AioContext polling feature but didn't enable it
> by default yet.  I have gone back over the performance data on the
> mailing list and picked a default polling value that gave good results.
> 
> Let's enable AioContext polling by default so users don't have another
> switch they need to set manually.  If performance regressions are found
> we can still disable this for the QEMU 2.9 release.
> 
> Cc: Paolo Bonzini <pbonzini@redhat.com>
> Cc: Christian Borntraeger <borntraeger@de.ibm.com>
> Cc: Karl Rister <krister@redhat.com>
> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> ---
>  iothread.c | 14 ++++++++++++++
>  1 file changed, 14 insertions(+)

Thanks, applied to my block tree:
https://github.com/stefanha/qemu/commits/block

Stefan

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 455 bytes --]