From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-9.8 required=3.0 tests=BAYES_00,DKIMWL_WL_HIGH, DKIM_SIGNED,DKIM_VALID,DKIM_VALID_AU,HEADER_FROM_DIFFERENT_DOMAINS, HTML_MESSAGE,INCLUDES_PATCH,MAILING_LIST_MULTI,NICE_REPLY_A,SPF_HELO_NONE, SPF_PASS,USER_AGENT_SANE_1 autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id A1C65C433DB for ; Wed, 6 Jan 2021 03:39:25 +0000 (UTC) Received: from kanga.kvack.org (kanga.kvack.org [205.233.56.17]) by mail.kernel.org (Postfix) with ESMTP id 1A17722D04 for ; Wed, 6 Jan 2021 03:39:24 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 1A17722D04 Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=oracle.com Authentication-Results: mail.kernel.org; spf=pass smtp.mailfrom=owner-linux-mm@kvack.org Received: by kanga.kvack.org (Postfix) id 069BF8D00DC; Tue, 5 Jan 2021 22:39:24 -0500 (EST) Received: by kanga.kvack.org (Postfix, from userid 40) id 019C98D00D1; Tue, 5 Jan 2021 22:39:23 -0500 (EST) X-Delivered-To: int-list-linux-mm@kvack.org Received: by kanga.kvack.org (Postfix, from userid 63042) id DAD948D00DC; Tue, 5 Jan 2021 22:39:23 -0500 (EST) X-Delivered-To: linux-mm@kvack.org Received: from forelay.hostedemail.com (smtprelay0092.hostedemail.com [216.40.44.92]) by kanga.kvack.org (Postfix) with ESMTP id C02538D00D1 for ; Tue, 5 Jan 2021 22:39:23 -0500 (EST) Received: from smtpin04.hostedemail.com (10.5.19.251.rfc1918.com [10.5.19.251]) by forelay04.hostedemail.com (Postfix) with ESMTP id 88DE71F08 for ; Wed, 6 Jan 2021 03:39:23 +0000 (UTC) X-FDA: 77673945006.04.tramp21_3305eb7274de Received: from filter.hostedemail.com (10.5.16.251.rfc1918.com [10.5.16.251]) by smtpin04.hostedemail.com (Postfix) with ESMTP id 704BB800F5B8 for ; Wed, 6 Jan 2021 03:39:23 +0000 (UTC) X-HE-Tag: tramp21_3305eb7274de X-Filterd-Recvd-Size: 22668 Received: from aserp2120.oracle.com (aserp2120.oracle.com [141.146.126.78]) by imf15.hostedemail.com (Postfix) with ESMTP for ; Wed, 6 Jan 2021 03:39:22 +0000 (UTC) Received: from pps.filterd (aserp2120.oracle.com [127.0.0.1]) by aserp2120.oracle.com (8.16.0.42/8.16.0.42) with SMTP id 1063T1eF195126; Wed, 6 Jan 2021 03:39:17 GMT DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=oracle.com; h=subject : to : cc : references : from : message-id : date : mime-version : in-reply-to : content-type; s=corp-2020-01-29; bh=D78+A7k7UHWgzEQXEQfyOqP0ZJE1YtZ2F3U3WszrsPo=; b=ZIm1T2IcaFSQxwNwFHoDzuVlNM7gVM+aui5nI5/XayUM8X6QblNFqpLhCwZgCyhaqx37 tp8EDoDtYWH/mW+FfO2e6qvJ04Z5udfSGiFtWZtI7fiD/d0PuRm/kFdEMhtTeX41iclm 2FUg0Yy8rqkDovOzxhpQUGxN+c9rR2/tYJwLH1YdnZsORlP7nXnDjrXvHg5aS+VvEQmW iZ7QM7S8ibuY3W2b2ScoogsyJFXQy0zF3MOqZ3MmfWkmD5iaMPmnoMBFW4lhLq8tu0xU yNKQz6E8eOQHvsMZNTNh74sk1EqU0zDGnBSoSkUTVIZu1fvIGryYhXP6uK49mAQcIjD3 YQ== Received: from aserp3020.oracle.com (aserp3020.oracle.com [141.146.126.70]) by aserp2120.oracle.com with ESMTP id 35w54201ct-1 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-GCM-SHA384 bits=256 verify=FAIL); Wed, 06 Jan 2021 03:39:16 +0000 Received: from pps.filterd (aserp3020.oracle.com [127.0.0.1]) by aserp3020.oracle.com (8.16.0.42/8.16.0.42) with SMTP id 1063cCx2154082; Wed, 6 Jan 2021 03:39:16 GMT Received: from userv0122.oracle.com (userv0122.oracle.com [156.151.31.75]) by aserp3020.oracle.com with ESMTP id 35v1f9e4n9-1 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-GCM-SHA384 bits=256 verify=OK); Wed, 06 Jan 2021 03:39:16 +0000 Received: from abhmp0017.oracle.com (abhmp0017.oracle.com [141.146.116.23]) by userv0122.oracle.com (8.14.4/8.14.4) with ESMTP id 1063dDfO030942; Wed, 6 Jan 2021 03:39:13 GMT Received: from [10.191.128.41] (/10.191.128.41) by default (Oracle Beehive Gateway v4.0) with ESMTP ; Wed, 06 Jan 2021 03:39:12 +0000 Subject: Re: [RFC PATCH] mm/memcontrol: Increase threshold for draining per-cpu stocked bytes. To: Roman Gushchin Cc: hannes@cmpxchg.org, mhocko@kernel.org, akpm@linux-foundation.org, cgroups@vger.kernel.org, linux-mm@kvack.org, linux-kernel@vger.kernel.org References: <1609862862-3573-1-git-send-email-imran.f.khan@oracle.com> <20210105182352.GE371241@carbon.dhcp.thefacebook.com> <20210105184558.GF371241@carbon.dhcp.thefacebook.com> <127fa24c-d4c4-5c24-ec30-ea6349f37923@oracle.com> <20210106032909.GG371241@carbon.dhcp.thefacebook.com> From: Imran Khan Message-ID: <9bd3018e-7182-fe4a-3ba2-ed0cf2e0875a@oracle.com> Date: Wed, 6 Jan 2021 14:39:05 +1100 User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:78.0) Gecko/20100101 Thunderbird/78.6.0 MIME-Version: 1.0 In-Reply-To: <20210106032909.GG371241@carbon.dhcp.thefacebook.com> Content-Type: multipart/alternative; boundary="------------CDBD7EDB8BDE9D89BA1373AC" Content-Language: en-US X-Proofpoint-Virus-Version: vendor=nai engine=6000 definitions=9855 signatures=668683 X-Proofpoint-Spam-Details: rule=notspam policy=default score=0 mlxlogscore=999 phishscore=0 suspectscore=0 spamscore=0 bulkscore=0 adultscore=0 mlxscore=0 malwarescore=0 classifier=spam adjust=0 reason=mlx scancount=1 engine=8.12.0-2009150000 definitions=main-2101060020 X-Proofpoint-Virus-Version: vendor=nai engine=6000 definitions=9855 signatures=668683 X-Proofpoint-Spam-Details: rule=notspam policy=default score=0 mlxscore=0 phishscore=0 adultscore=0 malwarescore=0 suspectscore=0 bulkscore=0 clxscore=1015 impostorscore=0 spamscore=0 mlxlogscore=999 lowpriorityscore=0 priorityscore=1501 classifier=spam adjust=0 reason=mlx scancount=1 engine=8.12.0-2009150000 definitions=main-2101060020 X-Bogosity: Ham, tests=bogofilter, spamicity=0.000000, version=1.2.4 Sender: owner-linux-mm@kvack.org Precedence: bulk X-Loop: owner-majordomo@kvack.org List-ID: This is a multi-part message in MIME format. --------------CDBD7EDB8BDE9D89BA1373AC Content-Type: text/plain; charset=utf-8; format=flowed Content-Transfer-Encoding: 7bit On 6/1/21 2:29 pm, Roman Gushchin wrote: > On Wed, Jan 06, 2021 at 02:07:12PM +1100, Imran Khan wrote: >> On 6/1/21 5:45 am, Roman Gushchin wrote: >>> On Tue, Jan 05, 2021 at 10:23:52AM -0800, Roman Gushchin wrote: >>>> On Tue, Jan 05, 2021 at 04:07:42PM +0000, Imran Khan wrote: >>>>> While allocating objects whose size is multiple of PAGE_SIZE, >>>>> say kmalloc-4K, we charge one page for extra bytes corresponding >>>>> to the obj_cgroup membership pointer and remainder of the charged >>>>> page gets added to per-cpu stocked bytes. If this allocation is >>>>> followed by another allocation of the same size, the stocked bytes >>>>> will not suffice and thus we endup charging an extra page >>>>> again for membership pointer and remainder of this page gets added >>>>> to per-cpu stocked bytes. This second addition will cause amount of >>>>> stocked bytes to go beyond PAGE_SIZE and hence will result in >>>>> invocation of drain_obj_stock. >>>>> >>>>> So if we are in a scenario where we are consecutively allocating, >>>>> several PAGE_SIZE multiple sized objects, the stocked bytes will >>>>> never be enough to suffice a request and every second request will >>>>> trigger draining of stocked bytes. >>>>> >>>>> For example invoking __alloc_skb multiple times with >>>>> 2K < packet size < 4K will give a call graph like: >>>>> >>>>> __alloc_skb >>>>> | >>>>> |__kmalloc_reserve.isra.61 >>>>> | | >>>>> | |__kmalloc_node_track_caller >>>>> | | | >>>>> | | |slab_pre_alloc_hook.constprop.88 >>>>> | | obj_cgroup_charge >>>>> | | | | >>>>> | | | |__memcg_kmem_charge >>>>> | | | | | >>>>> | | | | |page_counter_try_charge >>>>> | | | | >>>>> | | | |refill_obj_stock >>>>> | | | | | >>>>> | | | | |drain_obj_stock.isra.68 >>>>> | | | | | | >>>>> | | | | | |__memcg_kmem_uncharge >>>>> | | | | | | | >>>>> | | | | | | |page_counter_uncharge >>>>> | | | | | | | | >>>>> | | | | | | | |page_counter_cancel >>>>> | | | >>>>> | | | >>>>> | | |__slab_alloc >>>>> | | | | >>>>> | | | |___slab_alloc >>>>> | | | | >>>>> | | |slab_post_alloc_hook >>>>> >>>>> This frequent draining of stock bytes and resultant charging of pages >>>>> increases the CPU load and hence deteriorates the scheduler latency. >>>>> >>>>> The above mentioned scenario and it's impact can be seen by running >>>>> hackbench with large packet size on v5.8 and subsequent kernels. The >>>>> deterioration in hackbench number starts appearing from v5.9 kernel, >>>>> 'commit f2fe7b09a52b ("mm: memcg/slab: charge individual slab objects >>>>> instead of pages")'. >>>>> >>>>> Increasing the draining limit to twice of KMALLOC_MAX_CACHE_SIZE >>>>> (a safe upper limit for size of slab cache objects), will avoid draining >>>>> of stock, every second allocation request, for the above mentioned >>>>> scenario and hence will reduce the CPU load for such cases. For >>>>> allocation of smaller objects or other allocation patterns the behaviour >>>>> will be same as before. >>>>> >>>>> This change increases the draining threshold for per-cpu stocked bytes >>>>> from PAGE_SIZE to KMALLOC_MAX_CACHE_SIZE * 2. >>>> Hello, Imran! >>>> >>>> Yes, it makes total sense to me. >> Hi Roman, >> >> Thanks for reviewing this patch. >> >>>> Btw, in earlier versions of the new slab controller there was a separate stock >>>> for byte-sized charging and it was 32 pages large. Later Johannes suggested >>>> the current layered design and he thought that because of the layering a single >>>> page is enough for the upper layer. >>>> >>>>> Below are the hackbench numbers with and without this change on >>>>> v5.10.0-rc7. >>>>> >>>>> Without this change: >>>>> # hackbench process 10 1000 -s 100000 >>>>> Running in process mode with 10 groups using 40 file descriptors >>>>> each (== 400 tasks) >>>>> Each sender will pass 100 messages of 100000 bytes >>>>> Time: 4.401 >>>>> >>>>> # hackbench process 10 1000 -s 100000 >>>>> Running in process mode with 10 groups using 40 file descriptors >>>>> each (== 400 tasks) >>>>> Each sender will pass 100 messages of 100000 bytes >>>>> Time: 4.470 >>>>> >>>>> With this change: >>>>> # hackbench process 10 1000 -s 100000 >>>>> Running in process mode with 10 groups using 40 file descriptors >>>>> each (== 400 tasks) >>>>> Each sender will pass 100 messages of 100000 bytes >>>>> Time: 3.782 >>>>> >>>>> # hackbench process 10 1000 -s 100000 >>>>> Running in process mode with 10 groups using 40 file descriptors >>>>> each (== 400 tasks) >>>>> Each sender will pass 100 messages of 100000 bytes >>>>> Time: 3.827 >>>>> >>>>> As can be seen the change gives an improvement of about 15% in hackbench >>>>> numbers. >>>>> Also numbers obtained with the change are inline with those obtained >>>>> from v5.8 kernel. >>>> The difference is quite impressive! >>>> >>>> I wonder if you tried smaller values than KMALLOC_MAX_CACHE_SIZE * 2? >>>> Let's say 16 and 32? >> I have tested my change with smaller sizes as well and could see similar difference >> in hackbench numbers >> >> Without change(5.10.0-rc7 vanilla): >> >> # hackbench process 10 1000 -s 16 >> Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks) >> Each sender will pass 100 messages of 16 bytes >> Time: 0.429 >> >> # hackbench process 10 1000 -s 32 >> Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks) >> Each sender will pass 100 messages of 32 bytes >> Time: 0.458 >> >> With my changes on top of 5.10.0-rc7 >> # hackbench process 10 1000 -s 16 >> Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks) >> Each sender will pass 100 messages of 16 bytes >> Time: 0.347 >> >> # hackbench process 10 1000 -s 32 >> Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks) >> Each sender will pass 100 messages of 32 bytes >> Time: 0.324 >> >> I am confirming using BCC based argdist tool that these sizes result in call to >> __alloc_skb with size as 16 and 32 respectively. >> >>>> KMALLOC_MAX_CACHE_SIZE * 2 makes sense to me, but then the whole construction >>>> with two layer caching is very questionable. Anyway, it's not a reason to not >>>> merge your patch, just something I wanna look at later. >>> Hm, can you, please, benchmark the following change (without your change)? >>> >>> @@ -3204,7 +3204,7 @@ static void drain_obj_stock(struct memcg_stock_pcp *stock) >>> if (nr_pages) { >>> rcu_read_lock(); >>> - __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages); >>> + refill_stock(obj_cgroup_memcg(old), nr_pages); >>> rcu_read_unlock(); >>> } >> I have tested this change on top of v5.10-rc7 and this too gives performance improvement. >> I further confirmed using flamegraphs that with this change too we are avoiding following >> CPU intensive path >> >> |__memcg_kmem_uncharge >> | >> |page_counter_uncharge >> | | >> | |page_counter_cancel >> >> Please find the hackbench numbers with your change as given below: >> # hackbench process 10 1000 -s 100000 >> Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks) >> Each sender will pass 100 messages of 100000 bytes >> Time: 3.841 >> >> # hackbench process 10 1000 -s 100000 >> Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks) >> Each sender will pass 100 messages of 100000 bytes >> Time: 3.863 >> >> # hackbench process 10 1000 -s 16 >> Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks) >> Each sender will pass 100 messages of 16 bytes >> Time: 0.306 >> >> # hackbench process 10 1000 -s 32 >> Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks) >> Each sender will pass 100 messages of 32 bytes >> Time: 0.320 > Thank you for testing it! > > If there is no significant difference, I'd prefer to stick with this change instead of increasing > the size of the percpu batch, because it will preserve the accuracy of accounting. > > Will it work for you? Yes, this works for me too. Thanks, Imran > > Thanks! --------------CDBD7EDB8BDE9D89BA1373AC Content-Type: text/html; charset=utf-8 Content-Transfer-Encoding: 7bit 
On 6/1/21 2:29 pm, Roman Gushchin wrote:
On Wed, Jan 06, 2021 at 02:07:12PM +1100, Imran Khan wrote:

On 6/1/21 5:45 am, Roman Gushchin wrote:

On Tue, Jan 05, 2021 at 10:23:52AM -0800, Roman Gushchin wrote:

On Tue, Jan 05, 2021 at 04:07:42PM +0000, Imran Khan wrote:

While allocating objects whose size is multiple of PAGE_SIZE,
say kmalloc-4K, we charge one page for extra bytes corresponding
to the obj_cgroup membership pointer and remainder of the charged
page gets added to per-cpu stocked bytes. If this allocation is
followed by another allocation of the same size, the stocked bytes
will not suffice and thus we endup charging an extra page
again for membership pointer and remainder of this page gets added
to per-cpu stocked bytes. This second addition will cause amount of
stocked bytes to go beyond PAGE_SIZE and hence will result in
invocation of drain_obj_stock.

So if we are in a scenario where we are consecutively allocating,
several PAGE_SIZE multiple sized objects, the stocked bytes will
never be enough to suffice a request and every second request will
trigger draining of stocked bytes.

For example invoking __alloc_skb multiple times with
2K < packet size < 4K will give a call graph like:

__alloc_skb
     |
     |__kmalloc_reserve.isra.61
     |    |
     |    |__kmalloc_node_track_caller
     |    |    |
     |    |    |slab_pre_alloc_hook.constprop.88
     |    |     obj_cgroup_charge
     |    |    |    |
     |    |    |    |__memcg_kmem_charge
     |    |    |    |    |
     |    |    |    |    |page_counter_try_charge
     |    |    |    |
     |    |    |    |refill_obj_stock
     |    |    |    |    |
     |    |    |    |    |drain_obj_stock.isra.68
     |    |    |    |    |    |
     |    |    |    |    |    |__memcg_kmem_uncharge
     |    |    |    |    |    |    |
     |    |    |    |    |    |    |page_counter_uncharge
     |    |    |    |    |    |    |    |
     |    |    |    |    |    |    |    |page_counter_cancel
     |    |    |
     |    |    |
     |    |    |__slab_alloc
     |    |    |    |
     |    |    |    |___slab_alloc
     |    |    |    |
     |    |    |slab_post_alloc_hook

This frequent draining of stock bytes and resultant charging of pages
increases the CPU load and hence deteriorates the scheduler latency.

The above mentioned scenario and it's impact can be seen by running
hackbench with large packet size on v5.8 and subsequent kernels. The
deterioration in hackbench number starts appearing from v5.9 kernel,
'commit f2fe7b09a52b ("mm: memcg/slab: charge individual slab objects
instead of pages")'.

Increasing the draining limit to twice of KMALLOC_MAX_CACHE_SIZE
(a safe upper limit for size of slab cache objects), will avoid draining
of stock, every second allocation request, for the above mentioned
scenario and hence will reduce the CPU load for such cases. For
allocation of smaller objects or other allocation patterns the behaviour
will be same as before.

This change increases the draining threshold for per-cpu stocked bytes
from PAGE_SIZE to KMALLOC_MAX_CACHE_SIZE * 2.

Hello, Imran!

Yes, it makes total sense to me.

Hi Roman,

Thanks for reviewing this patch.


Btw, in earlier versions of the new slab controller there was a separate stock
for byte-sized charging and it was 32 pages large. Later Johannes suggested
the current layered design and he thought that because of the layering a single
page is enough for the upper layer.


Below are the hackbench numbers with and without this change on
v5.10.0-rc7.

Without this change:
     # hackbench process 10 1000 -s 100000
     Running in process mode with 10 groups using 40 file descriptors
     each (== 400 tasks)
     Each sender will pass 100 messages of 100000 bytes
     Time: 4.401

     # hackbench process 10 1000 -s 100000
     Running in process mode with 10 groups using 40 file descriptors
     each (== 400 tasks)
     Each sender will pass 100 messages of 100000 bytes
     Time: 4.470

With this change:
     # hackbench process 10 1000 -s 100000
     Running in process mode with 10 groups using 40 file descriptors
     each (== 400 tasks)
     Each sender will pass 100 messages of 100000 bytes
     Time: 3.782

     # hackbench process 10 1000 -s 100000
     Running in process mode with 10 groups using 40 file descriptors
     each (== 400 tasks)
     Each sender will pass 100 messages of 100000 bytes
     Time: 3.827

As can be seen the change gives an improvement of about 15% in hackbench
numbers.
Also numbers obtained with the change are inline with those obtained
from v5.8 kernel.

The difference is quite impressive!

I wonder if you tried smaller values than KMALLOC_MAX_CACHE_SIZE * 2?
Let's say 16 and 32?

I have tested my change with smaller sizes as well and could see similar difference
in hackbench numbers

Without change(5.10.0-rc7 vanilla):

# hackbench process 10 1000 -s 16
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 16 bytes
Time: 0.429

# hackbench process 10 1000 -s 32
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 32 bytes
Time: 0.458

With my changes on top of 5.10.0-rc7
# hackbench process 10 1000 -s 16
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 16 bytes
Time: 0.347

# hackbench process 10 1000 -s 32
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 32 bytes
Time: 0.324

I am confirming using BCC based argdist tool that these sizes result in call to
__alloc_skb with size as 16 and 32 respectively.


KMALLOC_MAX_CACHE_SIZE * 2 makes sense to me, but then the whole construction
with two layer caching is very questionable. Anyway, it's not a reason to not
merge your patch, just something I wanna look at later.

Hm, can you, please, benchmark the following change (without your change)?

@@ -3204,7 +3204,7 @@ static void drain_obj_stock(struct memcg_stock_pcp *stock)
  		if (nr_pages) {
  			rcu_read_lock();
-			__memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
+			refill_stock(obj_cgroup_memcg(old), nr_pages);
  			rcu_read_unlock();
  		}

I have tested this change on top of v5.10-rc7 and this too gives performance improvement.
I further confirmed using flamegraphs that with this change too we are avoiding following
CPU intensive path

|__memcg_kmem_uncharge
    |
    |page_counter_uncharge
    |    |
    |    |page_counter_cancel

Please find the hackbench numbers with your change as given below:
# hackbench process 10 1000 -s 100000
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 100000 bytes
Time: 3.841

# hackbench process 10 1000 -s 100000
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 100000 bytes
Time: 3.863

# hackbench process 10 1000 -s 16
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 16 bytes
Time: 0.306

# hackbench process 10 1000 -s 32
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 32 bytes
Time: 0.320

Thank you for testing it!

If there is no significant difference, I'd prefer to stick with this change instead of increasing
the size of the percpu batch, because it will preserve the accuracy of accounting.

Will it work for you?
Yes, this works for me too. 

Thanks,
Imran


Thanks!
--------------CDBD7EDB8BDE9D89BA1373AC--