persistent disk write at rate of ~4 MB/s

[Timur Irmatov <ir...@gmail.com>]

Hi,

I see that my trafficserver writes to disk at a rate of around 4 MB/s
irregardless of client load (which is small, less than 50 Mbit/s, with
as low as 2 Mbit/s at nights). I observe it with iostat and Collectd/
Graphite, 24 hours. Why is it doing so?

I am using bcache with one SSD and four 1 TB SATA drives behind it. I
am concerned that this persistent write stream will lead to premature
wear of SSD cache. Can I do something about it?

I think my question should be in FAQ. I noticed this behaviour a
couple of years ago but never got to asking about it until now.


-- 
Timur Irmatov

Re: persistent disk write at rate of ~4 MB/s

[Yongming Zhao <mi...@gmail.com>]

I don’t how bcache works with  the dir index sync writing, but I think those writes should not be read until next server restart, hopes bcache is smart enough to pass those write through.

if you want, you may try the interim caching, which works inside of ATS.


FYI


> 在 2014年10月19日，下午11:01，Leif Hedstrom <zw...@apache.org> 写道：
> 
> That sounds like the dir sync task. You can try changing the config to sync less frequently and see if that affects it? On very large disks it probably has to get very long, basically the dir sync throttles itself to not put too much pressure on the disks. Maybe we should make that configurable too.
> 
> -- Leif 
> 
> 
> 
>> On Oct 19, 2014, at 2:55 AM, Timur Irmatov <ir...@gmail.com> wrote:
>> 
>> Hi,
>> 
>> I see that my trafficserver writes to disk at a rate of around 4 MB/s
>> irregardless of client load (which is small, less than 50 Mbit/s, with
>> as low as 2 Mbit/s at nights). I observe it with iostat and Collectd/
>> Graphite, 24 hours. Why is it doing so?
>> 
>> I am using bcache with one SSD and four 1 TB SATA drives behind it. I
>> am concerned that this persistent write stream will lead to premature
>> wear of SSD cache. Can I do something about it?
>> 
>> I think my question should be in FAQ. I noticed this behaviour a
>> couple of years ago but never got to asking about it until now.
>> 
>> 
>> -- 
>> Timur Irmatov

- Yongming Zhao 赵永明

Re: persistent disk write at rate of ~4 MB/s

[Leif Hedstrom <zw...@apache.org>]

> On Oct 19, 2014, at 12:47 PM, Timur Irmatov <ir...@gmail.com> wrote:
> 
> On Sun, Oct 19, 2014 at 8:01 PM, Leif Hedstrom <zw...@apache.org> wrote:
>> That sounds like the dir sync task. You can try changing the config to sync less frequently and see if that affects it? On very large disks it probably has to get very long, basically the dir sync throttles itself to not put too much pressure on the disks. Maybe we should make that configurable too.
> 
> What is dir sync task? Could you elaborate? I haven't found any
> descriptions of that nor configuration settings.

  //  # how often should the directory be synced (seconds)
  {RECT_CONFIG, "proxy.config.cache.dir.sync_frequency", RECD_INT, "60", RECU_DYNAMIC, RR_NULL, RECC_NULL, NULL, RECA_NULL}
  ,

> 


So, traffic_server keeps an in-memory lookup table of disk cache items. In order to guarantee consistency, this RAM table has to be synced to disk periodically. You can reduce write pressure by increasing this sync frequency; However, be aware that doing so increases the window where you can lose disk data. This is safe, just means you could lose more data from showing up as written to disk during restarts.

— Leif

Re: persistent disk write at rate of ~4 MB/s

[Timur Irmatov <ir...@gmail.com>]

On Sun, Oct 19, 2014 at 8:01 PM, Leif Hedstrom <zw...@apache.org> wrote:
> That sounds like the dir sync task. You can try changing the config to sync less frequently and see if that affects it? On very large disks it probably has to get very long, basically the dir sync throttles itself to not put too much pressure on the disks. Maybe we should make that configurable too.

What is dir sync task? Could you elaborate? I haven't found any
descriptions of that nor configuration settings.


-- 
Timur Irmatov

Re: persistent disk write at rate of ~4 MB/s

[Leif Hedstrom <zw...@apache.org>]

That sounds like the dir sync task. You can try changing the config to sync less frequently and see if that affects it? On very large disks it probably has to get very long, basically the dir sync throttles itself to not put too much pressure on the disks. Maybe we should make that configurable too.

-- Leif 



> On Oct 19, 2014, at 2:55 AM, Timur Irmatov <ir...@gmail.com> wrote:
> 
> Hi,
> 
> I see that my trafficserver writes to disk at a rate of around 4 MB/s
> irregardless of client load (which is small, less than 50 Mbit/s, with
> as low as 2 Mbit/s at nights). I observe it with iostat and Collectd/
> Graphite, 24 hours. Why is it doing so?
> 
> I am using bcache with one SSD and four 1 TB SATA drives behind it. I
> am concerned that this persistent write stream will lead to premature
> wear of SSD cache. Can I do something about it?
> 
> I think my question should be in FAQ. I noticed this behaviour a
> couple of years ago but never got to asking about it until now.
> 
> 
> -- 
> Timur Irmatov