Mailing-List: contact dev-help@stdcxx.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@stdcxx.apache.org
Received-SPF: pass (nike.apache.org: domain of msebor@gmail.com designates
 209.85.210.182 as permitted sender)
Message-ID: <504823CE.2020005@gmail.com>
Date: Wed, 05 Sep 2012 22:17:18 -0600
From: Martin Sebor <msebor@gmail.com>
User-Agent: Mozilla/5.0 (X11; Linux x86_64;
 rv:15.0) Gecko/20120827 Thunderbird/15.0
MIME-Version: 1.0
To: Stefan Teleman <stefan.teleman@gmail.com>
CC: Apache Standard C++ Library <dev@stdcxx.apache.org>
Subject: Re: STDCXX-1056 [was: Re: STDCXX forks]
References: <40394653-8FCC-4D04-A108-2C650AF8F95B@hates.ms>
 <CALdE9OAN1H-jzGadddU4biEp-Bh2U1YAkPDZdsi9T=E5bKtirg@mail.gmail.com>
 <CALdE9ODQBKYVW+fG1eNTke1RK41ny221DPy+EYWv0OBbg3DKgA@mail.gmail.com>
 <5045E764.9090607@hates.ms>
 <CALdE9OC1foqvmgUXQ+Cvmt5yR=b7gfJpQcdtTxJAWp1_e8EtqA@mail.gmail.com>
 <595887D2-6E42-4BC4-AF69-085AE4BA8A7D@hates.ms>
 <CALdE9OBvZzgzMMqV5noQ0LANKBVi7=4UOK8_w43G00_PCD-7ZA@mail.gmail.com>
 <5046BDC1.3020400@gmail.com>
 <CALdE9ODJeZdyVmvDucCfzneTxmV86_3w9FqC7rOe56-5zdM0Xw@mail.gmail.com>
 <50476748.9040301@gmail.com>
 <CALdE9OD4RyJFq1Gci00G0ay1L80M-127Tuve9Nvx4NY5GtvKNw@mail.gmail.com>
 <50479CA6.8010306@hates.ms>
 <CALdE9OCx_OLwPhTJGKR7AYSA84ECHfiwru0n_kwUhfg_Sb3Qqw@mail.gmail.com>
 <CALdE9OC0cD6UOZ+DMK5D+51Z02tbt9gv4F8PTwL_zCeR9tn1sA@mail.gmail.com>
 <50481097.2080307@gmail.com>
 <CALdE9OCPSs6pcpMFZ+6rq6O+BXJJ+E-QEC1jDSoOb4xn2FiN=g@mail.gmail.com>
 <50481DDC.6010906@gmail.com>
In-Reply-To: <50481DDC.6010906@gmail.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit

Incidentally, a benchmark that I would expect to be more
representative of real programs than the test we've been
using is one that creates a small number of named locales
up front and then uses each in a loop in each thread (as
opposed to creating and destroying it in each iteration
and thread). With this setup, the per-object mutex
implementation should be much faster than the one with
the per-class mutex because of the absence of contention.

The test has an option that exercises this setup:
   --shared-locale

Martin

On 09/05/2012 09:51 PM, Martin Sebor wrote:
> On 09/05/2012 09:03 PM, Stefan Teleman wrote:
>> On Wed, Sep 5, 2012 at 10:55 PM, Martin Sebor <msebor@gmail.com> wrote:
>>
>>> I suspect the difference is due to the overhead of the repeated
>>> initialization and destruction of the per-object mutex in the
>>> test. The test repeatedly creates (and discards) named locale
>>> objects.
>>>
>>> The per-class mutex is initialized just once in the process, no
>>> matter how many facet objects (how many distinct named locales)
>>> the test creates. But the per-object mutex must be created (and
>>> destroyed) for each named locale.
>>
>> Agreed.
>>
>> But: if the choice is between an implementation which [1] breaks ABI
>> and [2] performs 20% worse -- even in contrived test cases -- than
>> another implementation [2] which doesn't break ABI, and performs
>> better than the first one,  why would we even consider the first one?
>
> Breaking the ABI is not an option (unless we rev the version).
> But I'm not sure I understand what you think breaks the ABI.
> We don't need to add a new mutex -- we can use the __rw_facet
> member for the locking. Or did you mean something else?
>
> Martin
>
>>
>> --Stefan
>>
>