Hi Levente,

On Jan 27, 2019, at 5:40 PM, Levente Uzonyi <leves@caesar.elte.hu> wrote:

On Sun, 27 Jan 2019, Chris Muller wrote:

Hi,

Yes, the SqueakMap server image is one part of the dynamic, but I

think another is a bug in the trunk image.  I think the reason Tim is

not seeing 45 seconds before error is because the timeout setting of

the high-up client is not being passed all the way down to the

lowest-level layers -- e.g., from HTTPSocket --> WebClient -->

SocketStream --> Socket.  By the time it gets down to Socket which

does the actual work, it's operating on its own 30 second timeout.

I would expect subsecond reponse times. 30 seconds is just unacceptably

long.

Well, it depends on if, for example, you're in the middle of

Antarctica with a slow internet connection in an office with a fast

connection.  A 30 second timeout is just the maximum amount of time

the client will wait for the entire process before presenting a

debugger, that's all it can do.

We can be sure that Tim should get subsecond response times instead of

timeouts after 30 seconds.

Right, but timeout settings are a necessary tool sometimes, my point

was that we should fix client code in trunk to make timeouts work

properly.

Incidentally, 99% of SqueakMap requests ARE subsecond -- just go to

map.squeak.org and click around and see.  For the remaining 1% that

aren't, the issue is known and we're working on a new server to fix

that.

Great! That was my point: the image needs to be fixed.

It is a fixed amount of time, I *think* still between 30 and 45

seconds, that it takes the SqueakMap server to save its model after an

and so if in the meantime it can simply be made to wait 45s instead of

30s, then current SqueakMap will only be that occasional delay at

worst, instead of the annoying debugger we currently get.

I don't see why that would make a difference: the user would get a debugger anyway, but only 15 seconds later.

You would save seconds, not milliseconds by not downloading files again.

IIUC, you're saying we would save one hope in the "download" --

instead of client <--> alan <--> andreas, it would just be client <-->

alan.  Is that right?

No. If the client doesn't have the mcz in the package cache but nginx has

it in its cache, then we save the transfer of data between alan and

andreas.

Are alan and andreas co-located?

They are cloud servers in the same data center.

The file doesn't have to be read from the disk either.

I assume you mean "read from disk" on alan?  What about after it's

cached so many mcz's in RAM that its paging out to swap file?  To me,

wasing precious RAM (of any server) to cache old MCZ file contents

that no one will ever download (because they become old very quickly)

feels wasteful.  Dragster cars are wasteful too, but yes, they are

"faster"... on a dragstrip.  :)  I guess there'd have to be some kind

of application-specific smart management of the cache...

Nginx's proxy_cache can handle that all automatically. Also, we don't need a large cache. A small, memory-only cache would do it.

Levente, what about the trunk directory listing, can it cache that?

Sure.

That is the _#1 thing_ source.squeak.org is accessing and sending back

over, and over, and over again -- every time that MC progress box that

says, "Updating [repository name]".

Right, unless you update an older image.

If the client does have the mcz, then we save the complete file transfer.

I don't know what the speed between alan <---> andreas is, but I doubt

it's much slower than client <---> alan in most cases, so the savings

would seem to be minimal..?

The image wouldn't have to open a file, read its content from the disk and

send that through a socket.

By "the image" I assume you mean the SqueakSource server image.  But

opening the file takes very little time.  Original web-sites were

.html files, remember how fast those were?  Plus, filesystems "cache"

file contents into their own internal caches anyway...

Each file uses one external semaphore, each socket uses three. If you use a default image, there can be no more than 256 external semaphores which is ridiculous for a server, and it'll just grind to a halt when some load arrives. Every time the external semaphore table is full, a GC is triggered to try clear it up via the finalization process.

Reading a file into memory is slow, writing it to a socket is slow.

(Compared to nginx which uses sendfile to let the kernel handle that).

And Squeak can only use a single process to handle everything.

That’s configurable.  Alas because writing lock-free table growth is not easy the external semaphore table doesn’t grow automatically.  But the vm does allow its size to be specified in a value cached in the image header and read at startup (IIRC).  So we could easily have a 4K entry external semaphore table.

Yes, it still has to return back through alan but I assume alan does

not wait for a "full download" received from andreas before its

already pipeing back to the Squeak client.  If true, then it seems

like it only amounts to saving one hop, which would hardly be

noticeable over what we have now.

The goal of caching is not about saving a hop, but to avoid handling files in Squeak.

Nginx does that thing magnitudes faster than

Squeak.

The UX would not be magnitudes faster though, right?

Directly by letting nginx serving the file, no, but the server image would be less likely to get stalled (return 5xx responses).

But the caching scheme I described in this thread would make the UX a lot quicker too, because data would not have to be transferred when you already have it.

That would also let us save bandwidth by not downloading files already

sitting in the client's package cache.

How so?  Isn't the package-cache checked before hitting the server at

all?  It certainly should be.

No, it's not. Currently that's not possible, because different files can

have the same name. And currently we have no way to tell them apart.

No.  No two MCZ's may have the same name, certainly not withiin the

same repository, because MCRepository cannot support that.  So maybe

Not at the same time, but it's possible, and it just happened recently

with Chronology-ul.21.

It is perfectly possible that a client has a version in its package cache

with the same name as a different version on the server.

But we don't want to restrict what's possible in our software design

because of that.  That situation is already a headache anyway.  Same

name theoretically can come only from the same person (if we ensure

unique initials) and so this is avoidable / fixable by resaving one of

them as a different name...

It wasn't me who created the duplicate. If your suggestion had been in place, some images out there, including mine, would have been broken by the update process.

we need project subdirectories under package-cache to properly

simulate each cached Repository.  I had no idea we were neutering 90%

of the benefits of our package-cache because of this too, and just

sitting here, I can't help wonder whether this is why MCProxy doesn't

work properly either!

The primary purpose of a cache is to *check it first* to speed up

access to something, right?  What you say about package-cache sounds

I don't know. It wasn't me who designed it. :)

I meant ANY "cache".

https://en.wikipedia.org/wiki/Cache_(computing)

It still depends on the purpose of the cache. It's possible that package-cache is just a misnomer or it was just a plan to use it as a cache which hasn't happened yet.

For Monticello, package-cache's other use-case is when an

authentication issue occurs when trying to save to a HTTP repository.

At that point the Version object with the new ancestry was already

constructed in memory, so rather than worry about trying to "undo" all

that, it was simpler and better to save it to a package-cache, persist

it safely so the client can simply move forward from there (get access

to the HTTP and copy it or whatever).

The package-cache is also handy as a default repository and as an offline storage.

Levente

- Chris

really bad we should fix that, not surrender to it.

Yes, that should be fixed, but it needs changes on the server side.

What I always had in mind was to extend the repository listing with

hashes/uuids so that the client could figure out if it needs to download a

specific version. But care must be taken not to break the code for

non-ss repositories (e.g. simple directory listings).

Levente

- Chris