[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
2016-02-29 12:51, Panu Matilainen:
> On 02/24/2016 03:23 PM, Ananyev, Konstantin wrote:
> > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Panu Matilainen
> >> On 02/23/2016 07:35 AM, Xie, Huawei wrote:
> >>> On 2/22/2016 10:52 PM, Xie, Huawei wrote:
> On 2/4/2016 1:24 AM, Olivier MATZ wrote:
> > On 01/27/2016 02:56 PM, Panu Matilainen wrote:
> >> Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of
> >> the library ABI and should not be listed in the version map.
> >>
> >> I assume its inline for performance reasons, but then you lose the
> >> benefits of dynamic linking such as ability to fix bugs and/or improve
> >> itby just updating the library. Since the point of having a bulk API is
> >> to improve performance by reducing the number of calls required, does
> >> it
> >> really have to be inline? As in, have you actually measured the
> >> difference between inline and non-inline and decided its worth all the
> >> downsides?
> > Agree with Panu. It would be interesting to compare the performance
> > between inline and non inline to decide whether inlining it or not.
> Will update after i gathered more data. inline could show obvious
> performance difference in some cases.
> >>>
> >>> Panu and Oliver:
> >>> I write a simple benchmark. This benchmark run 10M rounds, in each round
> >>> 8 mbufs are allocated through bulk API, and then freed.
> >>> These are the CPU cycles measured(Intel(R) Xeon(R) CPU E5-2680 0 @
> >>> 2.70GHz, CPU isolated, timer interrupt disabled, rcu offloaded).
> >>> Btw, i have removed some exceptional data, the frequency of which is
> >>> like 1/10. Sometimes observed user usage suddenly disappeared, no clue
> >>> what happened.
> >>>
> >>> With 8 mbufs allocated, there is about 6% performance increase using
> >>> inline.
> >> [...]
> >>>
> >>> With 16 mbufs allocated, we could still observe obvious performance
> >>> difference, though only 1%-2%
> >>>
> >> [...]
> >>>
> >>> With 32/64 mbufs allocated, the deviation of the data itself would hide
> >>> the performance difference.
> >>> So we prefer using inline for performance.
> >>
> >> At least I was more after real-world performance in a real-world
> >> use-case rather than CPU cycles in a microbenchmark, we know function
> >> calls have a cost but the benefits tend to outweight the cons.
> >>
> >> Inline functions have their place and they're far less evil in project
> >> internal use, but in library public API they are BAD and should be ...
> >> well, not banned because there are exceptions to every rule, but highly
> >> discouraged.
> >
> > Why is that?
>
> For all the reasons static linking is bad, and what's worse it forces
> the static linking badness into dynamically linked builds.
>
> If there's a bug (security or otherwise) in a library, a distro wants to
> supply an updated package which fixes that bug and be done with it. But
> if that bug is in an inlined code, supplying an update is not enough,
> you also need to recompile everything using that code, and somehow
> inform customers possibly using that code that they need to not only
> update the library but to recompile their apps as well. That is
> precisely the reason distros go to great lenghts to avoid *any*
> statically linked apps and libs in the distro, completely regardless of
> the performance overhead.
>
> In addition, inlined code complicates ABI compatibility issues because
> some of the code is one the "wrong" side, and worse, it bypasses all the
> other ABI compatibility safeguards like soname and symbol versioning.
>
> Like said, inlined code is fine for internal consumption, but incredibly
> bad for public interfaces. And of course, the more complicated a
> function is, greater the potential of needing bugfixes.
>
> Mind you, none of this is magically specific to this particular
> function. Except in the sense that bulk operations offer a better way of
> performance improvements than just inlining everything.
>
> > As you can see right now we have all mbuf alloc/free routines as static
> > inline.
> > And I think we would like to keep it like that.
> > So why that particular function should be different?
>
> Because there's much less need to have it inlined since the function
> call overhead is "amortized" by the fact its doing bulk operations. "We
> always did it that way" is not a very good reason :)
>
> > After all that function is nothing more than a wrapper
> > around rte_mempool_get_bulk() unrolled by 4 loop {rte_pktmbuf_reset()}
> > So unless mempool get/put API would change, I can hardly see there could be
> > any ABI
> > breakages in future.
> > About 'real world' performance gain - it was a 'real world' performance
> > problem,
> > that we tried to solve by introducing that function:
> > http://dpdk.org/ml/archives/dev/2015-May/017633.html
> >
> > And according to the user feedback, it does help:
> >
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 02/24/2016 03:23 PM, Ananyev, Konstantin wrote: > Hi Panu, > >> -Original Message- >> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Panu Matilainen >> Sent: Wednesday, February 24, 2016 12:12 PM >> To: Xie, Huawei; Olivier MATZ; dev at dpdk.org >> Cc: dprovan at bivio.net >> Subject: Re: [dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk >> API >> >> On 02/23/2016 07:35 AM, Xie, Huawei wrote: >>> On 2/22/2016 10:52 PM, Xie, Huawei wrote: >>>> On 2/4/2016 1:24 AM, Olivier MATZ wrote: >>>>> Hi, >>>>> >>>>> On 01/27/2016 02:56 PM, Panu Matilainen wrote: >>>>>> Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of >>>>>> the library ABI and should not be listed in the version map. >>>>>> >>>>>> I assume its inline for performance reasons, but then you lose the >>>>>> benefits of dynamic linking such as ability to fix bugs and/or improve >>>>>> itby just updating the library. Since the point of having a bulk API is >>>>>> to improve performance by reducing the number of calls required, does it >>>>>> really have to be inline? As in, have you actually measured the >>>>>> difference between inline and non-inline and decided its worth all the >>>>>> downsides? >>>>> Agree with Panu. It would be interesting to compare the performance >>>>> between inline and non inline to decide whether inlining it or not. >>>> Will update after i gathered more data. inline could show obvious >>>> performance difference in some cases. >>> >>> Panu and Oliver: >>> I write a simple benchmark. This benchmark run 10M rounds, in each round >>> 8 mbufs are allocated through bulk API, and then freed. >>> These are the CPU cycles measured(Intel(R) Xeon(R) CPU E5-2680 0 @ >>> 2.70GHz, CPU isolated, timer interrupt disabled, rcu offloaded). >>> Btw, i have removed some exceptional data, the frequency of which is >>> like 1/10. Sometimes observed user usage suddenly disappeared, no clue >>> what happened. >>> >>> With 8 mbufs allocated, there is about 6% performance increase using inline. >> [...] >>> >>> With 16 mbufs allocated, we could still observe obvious performance >>> difference, though only 1%-2% >>> >> [...] >>> >>> With 32/64 mbufs allocated, the deviation of the data itself would hide >>> the performance difference. >>> So we prefer using inline for performance. >> >> At least I was more after real-world performance in a real-world >> use-case rather than CPU cycles in a microbenchmark, we know function >> calls have a cost but the benefits tend to outweight the cons. >> >> Inline functions have their place and they're far less evil in project >> internal use, but in library public API they are BAD and should be ... >> well, not banned because there are exceptions to every rule, but highly >> discouraged. > > Why is that? For all the reasons static linking is bad, and what's worse it forces the static linking badness into dynamically linked builds. If there's a bug (security or otherwise) in a library, a distro wants to supply an updated package which fixes that bug and be done with it. But if that bug is in an inlined code, supplying an update is not enough, you also need to recompile everything using that code, and somehow inform customers possibly using that code that they need to not only update the library but to recompile their apps as well. That is precisely the reason distros go to great lenghts to avoid *any* statically linked apps and libs in the distro, completely regardless of the performance overhead. In addition, inlined code complicates ABI compatibility issues because some of the code is one the "wrong" side, and worse, it bypasses all the other ABI compatibility safeguards like soname and symbol versioning. Like said, inlined code is fine for internal consumption, but incredibly bad for public interfaces. And of course, the more complicated a function is, greater the potential of needing bugfixes. Mind you, none of this is magically specific to this particular function. Except in the sense that bulk operations offer a better way of performance improvements than just inlining everything. > As you can see right now we have all mbuf alloc/free routines as static > inline. > And I think we would like to keep it like that. > So why that particular function should be different? Because there's much less need to have it inlined
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
Hi Huawei,
On 02/26/2016 10:07 AM, Xie, Huawei wrote:
> On 2/26/2016 4:56 PM, Olivier MATZ wrote:
>> test_one_pktmbuf(struct rte_mbuf *m)
>> {
>> /* same as before without the allocation/free */
>> }
>>
>> test_pkt_mbuf(void)
>> {
>> m = rte_pktmbuf_alloc(pool);
>> test_one_pktmbuf(m);
>> rte_pktmbuf_free(m);
>>
>> ret = rte_pktmbuf_alloc_bulk(pool, mtab, BULK_CNT)
>> for (i = 0; i < BULK_CNT; i++) {
>> m = mtab[i];
>> test_one_pktmbuf(m);
>> rte_pktmbuf_free(m);
>> }
>> }
>
> This is to test the functionality.
> Let us also have the case like the following?
> cycles_start = rte_get_timer_cycles();
> while(rounds--) {
>
> ret = rte_pktmbuf_alloc_bulk(pool, mtab, BULK_CNT)
> for (i = 0; i < BULK_CNT; i++) {
> m = mtab[i];
> /* some work if needed */
> rte_pktmbuf_free(m);
> }
> }
> cycles_end = rte_get_timer_cycles();
>
> to compare with
>cycles_start = rte_get_timer_cycles();
>while(rounds--) {
> for (i = 0; i < BULK_CNT; i++)
> mtab[i] = rte_pktmbuf_alloc(...);
>
> ret = rte_pktmbuf_alloc_bulk(pool, mtab, BULK_CNT)
> for (i = 0; i < BULK_CNT; i++) {
> m = mtab[i];
> /* some work if needed */
> rte_pktmbuf_free(m);
> }
> }
> cycles_end = rte_get_timer_cycles();
In my opinion, it's already quite obvious that the bulk allocation
will be faster than the non-bulk (and we already have some mempool
benchmarks showing it). So I would say that functional testing is
enough.
On the other hand, it would be good to see if some examples
applications could be updated to take advantage of the new API (as
you did for the librte_vhost).
What do you think?
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 02/23/2016 06:35 AM, Xie, Huawei wrote:
>>> Also, it would be nice to have a simple test function in
>>> app/test/test_mbuf.c. For instance, you could update
>>> test_one_pktmbuf() to take a mbuf pointer as a parameter and remove
>>> the mbuf allocation from the function. Then it could be called with
>>> a mbuf allocated with rte_pktmbuf_alloc() (like before) and with
>>> all the mbufs of rte_pktmbuf_alloc_bulk().
>
> Don't quite get you. Is it that we write two cases, one case allocate
> mbuf through rte_pktmbuf_alloc_bulk and one use rte_pktmbuf_alloc? It is
> good to have.
Yes, something like:
test_one_pktmbuf(struct rte_mbuf *m)
{
/* same as before without the allocation/free */
}
test_pkt_mbuf(void)
{
m = rte_pktmbuf_alloc(pool);
test_one_pktmbuf(m);
rte_pktmbuf_free(m);
ret = rte_pktmbuf_alloc_bulk(pool, mtab, BULK_CNT)
for (i = 0; i < BULK_CNT; i++) {
m = mtab[i];
test_one_pktmbuf(m);
rte_pktmbuf_free(m);
}
}
> I could do this after this patch.
Yes, please.
Thanks,
Olivier
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 02/26/2016 08:39 AM, Xie, Huawei wrote:
With 8 mbufs allocated, there is about 6% performance increase using
inline.
With 16 mbufs allocated, we could still observe obvious performance
difference, though only 1%-2%
>
> On 2/24/2016 9:23 PM, Ananyev, Konstantin wrote:
>> As you can see right now we have all mbuf alloc/free routines as static
>> inline.
>> And I think we would like to keep it like that.
>> So why that particular function should be different?
>> After all that function is nothing more than a wrapper
>> around rte_mempool_get_bulk() unrolled by 4 loop {rte_pktmbuf_reset()}
>> So unless mempool get/put API would change, I can hardly see there could be
>> any ABI
>> breakages in future.
>> About 'real world' performance gain - it was a 'real world' performance
>> problem,
>> that we tried to solve by introducing that function:
>> http://dpdk.org/ml/archives/dev/2015-May/017633.html
>>
>> And according to the user feedback, it does help:
>> http://dpdk.org/ml/archives/dev/2016-February/033203.html
For me, there's no doubt this function will help in real world use
cases. That's also true that today most (oh no, all) datapath mbuf
functions are inline. Although I understand Panu's point of view
about the use of inline functions, trying to de-inline some functions
of the mbuf API (and others APIs like mempool or ring) would require
a deep analysis first to check the performance impact. And I think there
would be an impact for most of them.
In this particular case, as the function does bulk allocations, it
probably tempers the cost of the function call, and that's why I
was curious of any comparison with/without inlining. But I'm not
sure having this only function as non-inline makes a lot of sense.
So:
Acked-by: Olivier Matz
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 2/26/2016 4:56 PM, Olivier MATZ wrote:
>
> On 02/23/2016 06:35 AM, Xie, Huawei wrote:
Also, it would be nice to have a simple test function in
app/test/test_mbuf.c. For instance, you could update
test_one_pktmbuf() to take a mbuf pointer as a parameter and remove
the mbuf allocation from the function. Then it could be called with
a mbuf allocated with rte_pktmbuf_alloc() (like before) and with
all the mbufs of rte_pktmbuf_alloc_bulk().
>> Don't quite get you. Is it that we write two cases, one case allocate
>> mbuf through rte_pktmbuf_alloc_bulk and one use rte_pktmbuf_alloc? It is
>> good to have.
> Yes, something like:
>
> test_one_pktmbuf(struct rte_mbuf *m)
> {
> /* same as before without the allocation/free */
> }
>
> test_pkt_mbuf(void)
> {
> m = rte_pktmbuf_alloc(pool);
> test_one_pktmbuf(m);
> rte_pktmbuf_free(m);
>
> ret = rte_pktmbuf_alloc_bulk(pool, mtab, BULK_CNT)
> for (i = 0; i < BULK_CNT; i++) {
> m = mtab[i];
> test_one_pktmbuf(m);
> rte_pktmbuf_free(m);
> }
> }
This is to test the functionality.
Let us also have the case like the following?
cycles_start = rte_get_timer_cycles();
while(rounds--) {
ret = rte_pktmbuf_alloc_bulk(pool, mtab, BULK_CNT)
for (i = 0; i < BULK_CNT; i++) {
m = mtab[i];
/* some work if needed */
rte_pktmbuf_free(m);
}
}
cycles_end = rte_get_timer_cycles();
to compare with
cycles_start = rte_get_timer_cycles();
while(rounds--) {
for (i = 0; i < BULK_CNT; i++)
mtab[i] = rte_pktmbuf_alloc(...);
ret = rte_pktmbuf_alloc_bulk(pool, mtab, BULK_CNT)
for (i = 0; i < BULK_CNT; i++) {
m = mtab[i];
/* some work if needed */
rte_pktmbuf_free(m);
}
}
cycles_end = rte_get_timer_cycles();
>> I could do this after this patch.
> Yes, please.
>
>
> Thanks,
> Olivier
>
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 2/24/2016 9:23 PM, Ananyev, Konstantin wrote:
> Hi Panu,
>
>> -Original Message-
>> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Panu Matilainen
>> Sent: Wednesday, February 24, 2016 12:12 PM
>> To: Xie, Huawei; Olivier MATZ; dev at dpdk.org
>> Cc: dprovan at bivio.net
>> Subject: Re: [dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk
>> API
>>
>> On 02/23/2016 07:35 AM, Xie, Huawei wrote:
>>> On 2/22/2016 10:52 PM, Xie, Huawei wrote:
>>>> On 2/4/2016 1:24 AM, Olivier MATZ wrote:
>>>>> Hi,
>>>>>
>>>>> On 01/27/2016 02:56 PM, Panu Matilainen wrote:
>>>>>> Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of
>>>>>> the library ABI and should not be listed in the version map.
>>>>>>
>>>>>> I assume its inline for performance reasons, but then you lose the
>>>>>> benefits of dynamic linking such as ability to fix bugs and/or improve
>>>>>> itby just updating the library. Since the point of having a bulk API is
>>>>>> to improve performance by reducing the number of calls required, does it
>>>>>> really have to be inline? As in, have you actually measured the
>>>>>> difference between inline and non-inline and decided its worth all the
>>>>>> downsides?
>>>>> Agree with Panu. It would be interesting to compare the performance
>>>>> between inline and non inline to decide whether inlining it or not.
>>>> Will update after i gathered more data. inline could show obvious
>>>> performance difference in some cases.
>>> Panu and Oliver:
>>> I write a simple benchmark. This benchmark run 10M rounds, in each round
>>> 8 mbufs are allocated through bulk API, and then freed.
>>> These are the CPU cycles measured(Intel(R) Xeon(R) CPU E5-2680 0 @
>>> 2.70GHz, CPU isolated, timer interrupt disabled, rcu offloaded).
>>> Btw, i have removed some exceptional data, the frequency of which is
>>> like 1/10. Sometimes observed user usage suddenly disappeared, no clue
>>> what happened.
>>>
>>> With 8 mbufs allocated, there is about 6% performance increase using inline.
>> [...]
>>> With 16 mbufs allocated, we could still observe obvious performance
>>> difference, though only 1%-2%
>>>
>> [...]
>>> With 32/64 mbufs allocated, the deviation of the data itself would hide
>>> the performance difference.
>>> So we prefer using inline for performance.
>> At least I was more after real-world performance in a real-world
>> use-case rather than CPU cycles in a microbenchmark, we know function
>> calls have a cost but the benefits tend to outweight the cons.
It depends on what could be called the real world case. It could be
argued. I think the case Konstantin mentioned could be called a real
world one.
If your opinion on whether use benchmark or real-world use case is not
specific to this bulk API, then i have different opinion. For example,
for kernel virtio optimization, people use vring bench. We couldn't
guarantee each small optimization could bring obvious performance gain
in some big workload. The gain could be hided if bottleneck is
elsewhere, so i also plan to build such kind of virtio bench in DPDK.
Finally, i am open to inline or not, but currently priority better goes
with performance. If we make it an API now, we couldn't easily step back
in future; But we could change otherwise, after we have more confidence.
We could even check every inline "API", whether it should be inline or
be in the lib.
>>
>> Inline functions have their place and they're far less evil in project
>> internal use, but in library public API they are BAD and should be ...
>> well, not banned because there are exceptions to every rule, but highly
>> discouraged.
> Why is that?
> As you can see right now we have all mbuf alloc/free routines as static
> inline.
> And I think we would like to keep it like that.
> So why that particular function should be different?
> After all that function is nothing more than a wrapper
> around rte_mempool_get_bulk() unrolled by 4 loop {rte_pktmbuf_reset()}
> So unless mempool get/put API would change, I can hardly see there could be
> any ABI
> breakages in future.
> About 'real world' performance gain - it was a 'real world' performance
> problem,
> that we tried to solve by introducing that function:
> http://dpdk.org/ml/archives/dev/2015-May/017633.html
>
> And according to the user feedback, it does help:
> http://dpdk.org/ml/archives/dev/2016-February/033203.html
>
> Konstantin
>
>> - Panu -
>>
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 02/23/2016 07:35 AM, Xie, Huawei wrote: > On 2/22/2016 10:52 PM, Xie, Huawei wrote: >> On 2/4/2016 1:24 AM, Olivier MATZ wrote: >>> Hi, >>> >>> On 01/27/2016 02:56 PM, Panu Matilainen wrote: Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of the library ABI and should not be listed in the version map. I assume its inline for performance reasons, but then you lose the benefits of dynamic linking such as ability to fix bugs and/or improve itby just updating the library. Since the point of having a bulk API is to improve performance by reducing the number of calls required, does it really have to be inline? As in, have you actually measured the difference between inline and non-inline and decided its worth all the downsides? >>> Agree with Panu. It would be interesting to compare the performance >>> between inline and non inline to decide whether inlining it or not. >> Will update after i gathered more data. inline could show obvious >> performance difference in some cases. > > Panu and Oliver: > I write a simple benchmark. This benchmark run 10M rounds, in each round > 8 mbufs are allocated through bulk API, and then freed. > These are the CPU cycles measured(Intel(R) Xeon(R) CPU E5-2680 0 @ > 2.70GHz, CPU isolated, timer interrupt disabled, rcu offloaded). > Btw, i have removed some exceptional data, the frequency of which is > like 1/10. Sometimes observed user usage suddenly disappeared, no clue > what happened. > > With 8 mbufs allocated, there is about 6% performance increase using inline. [...] > > With 16 mbufs allocated, we could still observe obvious performance > difference, though only 1%-2% > [...] > > With 32/64 mbufs allocated, the deviation of the data itself would hide > the performance difference. > So we prefer using inline for performance. At least I was more after real-world performance in a real-world use-case rather than CPU cycles in a microbenchmark, we know function calls have a cost but the benefits tend to outweight the cons. Inline functions have their place and they're far less evil in project internal use, but in library public API they are BAD and should be ... well, not banned because there are exceptions to every rule, but highly discouraged. - Panu -
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
Hi Panu,
> -Original Message-
> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Panu Matilainen
> Sent: Wednesday, February 24, 2016 12:12 PM
> To: Xie, Huawei; Olivier MATZ; dev at dpdk.org
> Cc: dprovan at bivio.net
> Subject: Re: [dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk
> API
>
> On 02/23/2016 07:35 AM, Xie, Huawei wrote:
> > On 2/22/2016 10:52 PM, Xie, Huawei wrote:
> >> On 2/4/2016 1:24 AM, Olivier MATZ wrote:
> >>> Hi,
> >>>
> >>> On 01/27/2016 02:56 PM, Panu Matilainen wrote:
> >>>> Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of
> >>>> the library ABI and should not be listed in the version map.
> >>>>
> >>>> I assume its inline for performance reasons, but then you lose the
> >>>> benefits of dynamic linking such as ability to fix bugs and/or improve
> >>>> itby just updating the library. Since the point of having a bulk API is
> >>>> to improve performance by reducing the number of calls required, does it
> >>>> really have to be inline? As in, have you actually measured the
> >>>> difference between inline and non-inline and decided its worth all the
> >>>> downsides?
> >>> Agree with Panu. It would be interesting to compare the performance
> >>> between inline and non inline to decide whether inlining it or not.
> >> Will update after i gathered more data. inline could show obvious
> >> performance difference in some cases.
> >
> > Panu and Oliver:
> > I write a simple benchmark. This benchmark run 10M rounds, in each round
> > 8 mbufs are allocated through bulk API, and then freed.
> > These are the CPU cycles measured(Intel(R) Xeon(R) CPU E5-2680 0 @
> > 2.70GHz, CPU isolated, timer interrupt disabled, rcu offloaded).
> > Btw, i have removed some exceptional data, the frequency of which is
> > like 1/10. Sometimes observed user usage suddenly disappeared, no clue
> > what happened.
> >
> > With 8 mbufs allocated, there is about 6% performance increase using inline.
> [...]
> >
> > With 16 mbufs allocated, we could still observe obvious performance
> > difference, though only 1%-2%
> >
> [...]
> >
> > With 32/64 mbufs allocated, the deviation of the data itself would hide
> > the performance difference.
> > So we prefer using inline for performance.
>
> At least I was more after real-world performance in a real-world
> use-case rather than CPU cycles in a microbenchmark, we know function
> calls have a cost but the benefits tend to outweight the cons.
>
> Inline functions have their place and they're far less evil in project
> internal use, but in library public API they are BAD and should be ...
> well, not banned because there are exceptions to every rule, but highly
> discouraged.
Why is that?
As you can see right now we have all mbuf alloc/free routines as static inline.
And I think we would like to keep it like that.
So why that particular function should be different?
After all that function is nothing more than a wrapper
around rte_mempool_get_bulk() unrolled by 4 loop {rte_pktmbuf_reset()}
So unless mempool get/put API would change, I can hardly see there could be any
ABI
breakages in future.
About 'real world' performance gain - it was a 'real world' performance problem,
that we tried to solve by introducing that function:
http://dpdk.org/ml/archives/dev/2015-May/017633.html
And according to the user feedback, it does help:
http://dpdk.org/ml/archives/dev/2016-February/033203.html
Konstantin
>
> - Panu -
>
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 2/22/2016 10:52 PM, Xie, Huawei wrote: > On 2/4/2016 1:24 AM, Olivier MATZ wrote: >> Hi, >> >> On 01/27/2016 02:56 PM, Panu Matilainen wrote: >>> Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of >>> the library ABI and should not be listed in the version map. >>> >>> I assume its inline for performance reasons, but then you lose the >>> benefits of dynamic linking such as ability to fix bugs and/or improve >>> itby just updating the library. Since the point of having a bulk API is >>> to improve performance by reducing the number of calls required, does it >>> really have to be inline? As in, have you actually measured the >>> difference between inline and non-inline and decided its worth all the >>> downsides? >> Agree with Panu. It would be interesting to compare the performance >> between inline and non inline to decide whether inlining it or not. > Will update after i gathered more data. inline could show obvious > performance difference in some cases. Panu and Oliver: I write a simple benchmark. This benchmark run 10M rounds, in each round 8 mbufs are allocated through bulk API, and then freed. These are the CPU cycles measured(Intel(R) Xeon(R) CPU E5-2680 0 @ 2.70GHz, CPU isolated, timer interrupt disabled, rcu offloaded). Btw, i have removed some exceptional data, the frequency of which is like 1/10. Sometimes observed user usage suddenly disappeared, no clue what happened. With 8 mbufs allocated, there is about 6% performance increase using inline. inlinenon-inline 2780732950309416 28348536962951378072 28230153202954500888 28250600322958939912 28244998042898938284 28108597202944892796 28522294203014273296 27873085002956809852 27933372602958674900 2834762954346352 27854551842925719136 28215286242937380416 28229221362974978604 27766459202947666548 28159525722952316900 28010487402947366984 28514626722946469004 With 16 mbufs allocated, we could still observe obvious performance difference, though only 1%-2% inlinenon-inline 55199870845669902680 55384160965737646840 55789340645590165532 55481319725767926840 56255856965831345628 55582828765662223764 54455877685641003924 55590963205775258444 56564379885743969272 54409394045664882412 54988759685785138532 55616528085737123940 55152117165627775604 55505671405630790628 56659642805589568164 55912959005702697308 With 32/64 mbufs allocated, the deviation of the data itself would hide the performance difference. So we prefer using inline for performance. >> Also, it would be nice to have a simple test function in >> app/test/test_mbuf.c. For instance, you could update >> test_one_pktmbuf() to take a mbuf pointer as a parameter and remove >> the mbuf allocation from the function. Then it could be called with >> a mbuf allocated with rte_pktmbuf_alloc() (like before) and with >> all the mbufs of rte_pktmbuf_alloc_bulk(). Don't quite get you. Is it that we write two cases, one case allocate mbuf through rte_pktmbuf_alloc_bulk and one use rte_pktmbuf_alloc? It is good to have. I could do this after this patch. >> >> Regards, >> Olivier >> >
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 2/4/2016 1:24 AM, Olivier MATZ wrote: > Hi, > > On 01/27/2016 02:56 PM, Panu Matilainen wrote: >> >> Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of >> the library ABI and should not be listed in the version map. >> >> I assume its inline for performance reasons, but then you lose the >> benefits of dynamic linking such as ability to fix bugs and/or improve >> itby just updating the library. Since the point of having a bulk API is >> to improve performance by reducing the number of calls required, does it >> really have to be inline? As in, have you actually measured the >> difference between inline and non-inline and decided its worth all the >> downsides? > > Agree with Panu. It would be interesting to compare the performance > between inline and non inline to decide whether inlining it or not. Will update after i gathered more data. inline could show obvious performance difference in some cases. > > Also, it would be nice to have a simple test function in > app/test/test_mbuf.c. For instance, you could update > test_one_pktmbuf() to take a mbuf pointer as a parameter and remove > the mbuf allocation from the function. Then it could be called with > a mbuf allocated with rte_pktmbuf_alloc() (like before) and with > all the mbufs of rte_pktmbuf_alloc_bulk(). > > Regards, > Olivier >
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
Hi, On 01/27/2016 02:56 PM, Panu Matilainen wrote: > > Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of > the library ABI and should not be listed in the version map. > > I assume its inline for performance reasons, but then you lose the > benefits of dynamic linking such as ability to fix bugs and/or improve > itby just updating the library. Since the point of having a bulk API is > to improve performance by reducing the number of calls required, does it > really have to be inline? As in, have you actually measured the > difference between inline and non-inline and decided its worth all the > downsides? Agree with Panu. It would be interesting to compare the performance between inline and non inline to decide whether inlining it or not. Also, it would be nice to have a simple test function in app/test/test_mbuf.c. For instance, you could update test_one_pktmbuf() to take a mbuf pointer as a parameter and remove the mbuf allocation from the function. Then it could be called with a mbuf allocated with rte_pktmbuf_alloc() (like before) and with all the mbufs of rte_pktmbuf_alloc_bulk(). Regards, Olivier
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
On 01/26/2016 07:03 PM, Huawei Xie wrote:
> v6 changes:
> reflect the changes in release notes and library version map file
> revise our duff's code style a bit to make it more readable
>
> v5 changes:
> add comment about duff's device and our variant implementation
>
> v3 changes:
> move while after case 0
> add context about duff's device and why we use while loop in the commit
> message
>
> v2 changes:
> unroll the loop a bit to help the performance
>
> rte_pktmbuf_alloc_bulk allocates a bulk of packet mbufs.
>
> There is related thread about this bulk API.
> http://dpdk.org/dev/patchwork/patch/4718/
> Thanks to Konstantin's loop unrolling.
>
> Attached the wiki page about duff's device. It explains the performance
> optimization through loop unwinding, and also the most dramatic use of
> case label fall-through.
> https://en.wikipedia.org/wiki/Duff%27s_device
>
> In our implementation, we use while() loop rather than do{} while() loop
> because we could not assume count is strictly positive. Using while()
> loop saves one line of check if count is zero.
>
> Signed-off-by: Gerald Rogers
> Signed-off-by: Huawei Xie
> Acked-by: Konstantin Ananyev
> ---
> doc/guides/rel_notes/release_2_3.rst | 3 ++
> lib/librte_mbuf/rte_mbuf.h | 55
>
> lib/librte_mbuf/rte_mbuf_version.map | 7 +
> 3 files changed, 65 insertions(+)
>
> diff --git a/doc/guides/rel_notes/release_2_3.rst
> b/doc/guides/rel_notes/release_2_3.rst
> index 99de186..a52cba3 100644
> --- a/doc/guides/rel_notes/release_2_3.rst
> +++ b/doc/guides/rel_notes/release_2_3.rst
> @@ -4,6 +4,9 @@ DPDK Release 2.3
> New Features
>
>
> +* **Enable bulk allocation of mbufs. **
> + A new function ``rte_pktmbuf_alloc_bulk()`` has been added to allow the
> user
> + to allocate a bulk of mbufs.
>
> Resolved Issues
> ---
> diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
> index f234ac9..b2ed479 100644
> --- a/lib/librte_mbuf/rte_mbuf.h
> +++ b/lib/librte_mbuf/rte_mbuf.h
> @@ -1336,6 +1336,61 @@ static inline struct rte_mbuf
> *rte_pktmbuf_alloc(struct rte_mempool *mp)
> }
>
> /**
> + * Allocate a bulk of mbufs, initialize refcnt and reset the fields to
> default
> + * values.
> + *
> + * @param pool
> + *The mempool from which mbufs are allocated.
> + * @param mbufs
> + *Array of pointers to mbufs
> + * @param count
> + *Array size
> + * @return
> + * - 0: Success
> + */
> +static inline int rte_pktmbuf_alloc_bulk(struct rte_mempool *pool,
> + struct rte_mbuf **mbufs, unsigned count)
> +{
> + unsigned idx = 0;
> + int rc;
> +
> + rc = rte_mempool_get_bulk(pool, (void **)mbufs, count);
> + if (unlikely(rc))
> + return rc;
> +
> + /* To understand duff's device on loop unwinding optimization, see
> + * https://en.wikipedia.org/wiki/Duff's_device.
> + * Here while() loop is used rather than do() while{} to avoid extra
> + * check if count is zero.
> + */
> + switch (count % 4) {
> + case 0:
> + while (idx != count) {
> + RTE_MBUF_ASSERT(rte_mbuf_refcnt_read(mbufs[idx]) == 0);
> + rte_mbuf_refcnt_set(mbufs[idx], 1);
> + rte_pktmbuf_reset(mbufs[idx]);
> + idx++;
> + case 3:
> + RTE_MBUF_ASSERT(rte_mbuf_refcnt_read(mbufs[idx]) == 0);
> + rte_mbuf_refcnt_set(mbufs[idx], 1);
> + rte_pktmbuf_reset(mbufs[idx]);
> + idx++;
> + case 2:
> + RTE_MBUF_ASSERT(rte_mbuf_refcnt_read(mbufs[idx]) == 0);
> + rte_mbuf_refcnt_set(mbufs[idx], 1);
> + rte_pktmbuf_reset(mbufs[idx]);
> + idx++;
> + case 1:
> + RTE_MBUF_ASSERT(rte_mbuf_refcnt_read(mbufs[idx]) == 0);
> + rte_mbuf_refcnt_set(mbufs[idx], 1);
> + rte_pktmbuf_reset(mbufs[idx]);
> + idx++;
> + }
> + }
> + return 0;
> +}
> +
> +/**
>* Attach packet mbuf to another packet mbuf.
>*
>* After attachment we refer the mbuf we attached as 'indirect',
> diff --git a/lib/librte_mbuf/rte_mbuf_version.map
> b/lib/librte_mbuf/rte_mbuf_version.map
> index e10f6bd..257c65a 100644
> --- a/lib/librte_mbuf/rte_mbuf_version.map
> +++ b/lib/librte_mbuf/rte_mbuf_version.map
> @@ -18,3 +18,10 @@ DPDK_2.1 {
> rte_pktmbuf_pool_create;
>
> } DPDK_2.0;
> +
> +DPDK_2.3 {
> + global:
> +
> + rte_pktmbuf_alloc_bulk;
> +
> +} DPDK_2.1;
>
Since rte_pktmbuf_alloc_bulk() is an inline function, it is not part of
the library ABI and should not be listed in the version map.
I assume its inline for performance reasons, but then you lose the
benefits of dynamic linking such as ability to fix bugs and/or improve
itby just updating the library. Since the point of
[dpdk-dev] [PATCH v6 1/2] mbuf: provide rte_pktmbuf_alloc_bulk API
v6 changes:
reflect the changes in release notes and library version map file
revise our duff's code style a bit to make it more readable
v5 changes:
add comment about duff's device and our variant implementation
v3 changes:
move while after case 0
add context about duff's device and why we use while loop in the commit
message
v2 changes:
unroll the loop a bit to help the performance
rte_pktmbuf_alloc_bulk allocates a bulk of packet mbufs.
There is related thread about this bulk API.
http://dpdk.org/dev/patchwork/patch/4718/
Thanks to Konstantin's loop unrolling.
Attached the wiki page about duff's device. It explains the performance
optimization through loop unwinding, and also the most dramatic use of
case label fall-through.
https://en.wikipedia.org/wiki/Duff%27s_device
In our implementation, we use while() loop rather than do{} while() loop
because we could not assume count is strictly positive. Using while()
loop saves one line of check if count is zero.
Signed-off-by: Gerald Rogers
Signed-off-by: Huawei Xie
Acked-by: Konstantin Ananyev
---
doc/guides/rel_notes/release_2_3.rst | 3 ++
lib/librte_mbuf/rte_mbuf.h | 55
lib/librte_mbuf/rte_mbuf_version.map | 7 +
3 files changed, 65 insertions(+)
diff --git a/doc/guides/rel_notes/release_2_3.rst
b/doc/guides/rel_notes/release_2_3.rst
index 99de186..a52cba3 100644
--- a/doc/guides/rel_notes/release_2_3.rst
+++ b/doc/guides/rel_notes/release_2_3.rst
@@ -4,6 +4,9 @@ DPDK Release 2.3
New Features
+* **Enable bulk allocation of mbufs. **
+ A new function ``rte_pktmbuf_alloc_bulk()`` has been added to allow the user
+ to allocate a bulk of mbufs.
Resolved Issues
---
diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
index f234ac9..b2ed479 100644
--- a/lib/librte_mbuf/rte_mbuf.h
+++ b/lib/librte_mbuf/rte_mbuf.h
@@ -1336,6 +1336,61 @@ static inline struct rte_mbuf *rte_pktmbuf_alloc(struct
rte_mempool *mp)
}
/**
+ * Allocate a bulk of mbufs, initialize refcnt and reset the fields to default
+ * values.
+ *
+ * @param pool
+ *The mempool from which mbufs are allocated.
+ * @param mbufs
+ *Array of pointers to mbufs
+ * @param count
+ *Array size
+ * @return
+ * - 0: Success
+ */
+static inline int rte_pktmbuf_alloc_bulk(struct rte_mempool *pool,
+struct rte_mbuf **mbufs, unsigned count)
+{
+ unsigned idx = 0;
+ int rc;
+
+ rc = rte_mempool_get_bulk(pool, (void **)mbufs, count);
+ if (unlikely(rc))
+ return rc;
+
+ /* To understand duff's device on loop unwinding optimization, see
+* https://en.wikipedia.org/wiki/Duff's_device.
+* Here while() loop is used rather than do() while{} to avoid extra
+* check if count is zero.
+*/
+ switch (count % 4) {
+ case 0:
+ while (idx != count) {
+ RTE_MBUF_ASSERT(rte_mbuf_refcnt_read(mbufs[idx]) == 0);
+ rte_mbuf_refcnt_set(mbufs[idx], 1);
+ rte_pktmbuf_reset(mbufs[idx]);
+ idx++;
+ case 3:
+ RTE_MBUF_ASSERT(rte_mbuf_refcnt_read(mbufs[idx]) == 0);
+ rte_mbuf_refcnt_set(mbufs[idx], 1);
+ rte_pktmbuf_reset(mbufs[idx]);
+ idx++;
+ case 2:
+ RTE_MBUF_ASSERT(rte_mbuf_refcnt_read(mbufs[idx]) == 0);
+ rte_mbuf_refcnt_set(mbufs[idx], 1);
+ rte_pktmbuf_reset(mbufs[idx]);
+ idx++;
+ case 1:
+ RTE_MBUF_ASSERT(rte_mbuf_refcnt_read(mbufs[idx]) == 0);
+ rte_mbuf_refcnt_set(mbufs[idx], 1);
+ rte_pktmbuf_reset(mbufs[idx]);
+ idx++;
+ }
+ }
+ return 0;
+}
+
+/**
* Attach packet mbuf to another packet mbuf.
*
* After attachment we refer the mbuf we attached as 'indirect',
diff --git a/lib/librte_mbuf/rte_mbuf_version.map
b/lib/librte_mbuf/rte_mbuf_version.map
index e10f6bd..257c65a 100644
--- a/lib/librte_mbuf/rte_mbuf_version.map
+++ b/lib/librte_mbuf/rte_mbuf_version.map
@@ -18,3 +18,10 @@ DPDK_2.1 {
rte_pktmbuf_pool_create;
} DPDK_2.0;
+
+DPDK_2.3 {
+ global:
+
+ rte_pktmbuf_alloc_bulk;
+
+} DPDK_2.1;
--
1.8.1.4
