1 |
On 26/08/2020 18:40, Grant Taylor wrote: |
2 |
> On 8/21/20 10:15 PM, Caveman Al Toraboran wrote: |
3 |
>> just to double check i got you right. due to flushing the buffer to |
4 |
>> disk, this would mean that mail's throughput is limited by disk i/o? |
5 |
> |
6 |
> Yes. |
7 |
> |
8 |
> This speed limitation is viewed as a necessary limitation for the safety |
9 |
> of email passing through the system. |
10 |
> |
11 |
> Nothing states that it must be a single disk (block device). It's |
12 |
> entirely possible that a fancy MTA can rotate through many disks (block |
13 |
> devices), using a different one for each SMTP connection. Thus in |
14 |
> theory allowing some to operate in close lock step with each other |
15 |
> without depending on / being blocked by any given disk (block device). |
16 |
> |
17 |
> Thank you for the detailed explanation Ashley. |
18 |
|
19 |
Or think back to the old days - network was slow and disks were |
20 |
relatively fast. The disk was more than capable of keeping up with the |
21 |
network, and simple winchesters didn't lie about saving to the rotating |
22 |
rust ... |
23 |
> |
24 |
> As Ashley explained, some MTAs trust the kernel. I've heard of others |
25 |
> issuing a sync after the write. But that is up to each MTA's |
26 |
> developers. They have all taken reasonable steps to ensure the safety |
27 |
> of email. Some have taken more-than-reasonable steps. |
28 |
> |
29 |
Depends on the filesystem. "sync after write" was an EXTREMELY daft idea |
30 |
on ext4 in the early days - that would really have killed system response. |
31 |
|
32 |
Nowadays you could stick an SSD cache in front of a raid array ... |
33 |
|
34 |
Cheers, |
35 |
Wol |