Glaber modules, Go/Python support thoughts and research

The initial research and lookup on modules and Go/Python possibilities.

First research and googling have revealed that it’s pretty standard features – to extend C functionality via using Go libs and there is a way to include and execute Python code inside C.

Then I finally decided to look deeper at existing Zabbix loadable module functionality.

I’ve already come across them before, while I was digging history syncers for some problems. That time I thought I might be better extending the History interface by making modules but then come to the conclusion that they have very very limited functionality to have a fully functional history interface to be implemented via modules.

Which is strange.

Yes, because it’s only a small step – to make a history read interface and then we would have a modular interface which would allow only by attaching a proper module to achieve all that fast and fancy-schmancy features like ClickHouse integration or, perhaps, any other History Storage which fits well for historical time series data.

So this has definitely to be done now. Fortunately, it’s quite simple functionality and I stress this again – I don’t understand why it hasn’t been done right out of the box.

Second. The load_modules procedure. It’s quite obvious and simple, just a few things I need to do there. I’ll need to do some extra functionality to do different processing of modules due to module callbacks that will be different.

Sure, new API_VERSION which I’ll call GLABER API VERSION must be declared by the module to be treated and loaded differently from original Zabbix modules.

In initial modules, there has been some intermediate structure to hold pointers to history callbacks responsible for different history types.

I don’t understand the profit of it. Maybe yet. But there is a big disadvantage to it.

I will be extending the list of callbacks: history, polling, perhaps preprocessing functions might be extended soon by custom community code.

So the intermediate structure would have to be changed each time when a new callback type appears.

This way all the modules will have to be rebuilt from sources on such a change. No good. So far I haven’t found a reason why wouldn’t use some predefined function names and use dlsym to find them and attach to a list of callbacks of certain functionality. 

After digging all that and understanding Glaber code changes, I decided to start the initial implementation of the Go module. 

But before doing something it’s always worth of searching – it’s very unlikely no one has done that already. 

And that’s right: there is a nice place – Zabbix Share where there are plenty of modules and adaptors with examples how to code on Go, Python, PHP to make agent modules.

 It’s not a big difference since it’s the same structure and functionality for agents. And some of them exist for ages. For example, the Golang module is dated by 2015 !.

This a little bit sad. I say this again. If Zabbix had had full history interface implemented, which only literally needs a few days from them, then today Zabbix would work on any storage thing for years already. 

But ok, we can do it right now. It's never too late.

I have decided to start with Golang. It stays a bit apart from interpreted languages like Python/PHP. It’s compiled to a binary. It's faster. 

Since some point after Go 1.3-1.5, it’s possible to build modules without tricks with renaming the “main” function which was mandatory before. Fully compiled to a binary code Go modules do not need adapters and might be directly built to shared libraries.

A few notes about Go itself – after looking through for it's features I have realized it’s just a perfect for most things I’ve already done in Glaber code: It has the very nice concept of channels that perfectly fits tasks like streaming flow of metrics to history storage. It has a lightweight pseudo-threading named "goroutines" which makes perfectly easy coding of async polling – you can code it in simple sync way, and Goroutines solve the problem of thread efficiency.

Sure, that all must be tested and checked first. I have some doubts if all that nice Go features will function in library mode when the only certain function will be called from C code. But even if they will not, it’s still worth using Go for code simplification.

Then, memory leaks. It is a concern now. Since there will be lots of new code running, it is something that must be checked well. 

Especially in the case of Python. From my previous experience with a similar task with FreeRadius and rlm_perl it’s quite hard to eliminate memory leaks sometimes – too much code and functionality involved to be able to solve the thing in a reasonable time.

Ok, the actual tests. So far I have played a bit with compiling Go code to libraries and have tried to attach the libraries to a simple C code. Sometimes it has worked, sometimes it didn’t, but it’s enough to conclude that it’s perfectly doable and functional.

So the first plan will be to alter a bit Loading of Modules, implement support of history interface callbacks without touching existing modules support. And then after implementing basic reads and writes process a few millions of metrics, and look at my primary concerns about Golang:

1. Will go channels work?

2. Will we see any major memory leaks?

It’s most likely, that existing history_clickhouse will be moved to a separate module as well to compare C and Go efficiency if needed.

That’s it for now.

The “monitoring domains” idea

This one of the most important articles about the clustered Zabbix monitoring. It’s about common terminology and architecture.

When there is a classical server with a set of proxies, there is a strict host-> server or host -> proxy relation. This is done by setting the proxy attribute to a host in Zabbix configuration panels.

But when we have a cluster, then we would like to bound a host to a set of servers or proxies. And this is what “monitoring domains” for.

So, a “monitoring domain” is an identifier of set of proxies and servers which should be used to process a certain host.

A monitoring domain name may look just like an ordinary internet domain name “service.example.org” or you may use any kind of string which is valid for domain name like “server_farm1_major_hosts”, however the first one looks better.

Let’s look from a user’s perspective to “monitoring domains”.

Say, one got a server in the central office and two proxies in two branches.  

In a classic configuration this might be a server and two proxies.

Hosts from branch1 are bound to proxy1 just as hosts in the branch2 are set to be monitored from proxy in branch2. All hosts from central office doesn’t have a proxy set, so they are monitored directly by the server.

In a clustered architecture for such a case you would want to define three domains – central_office, branch1 and branch2. Since cluster is an option thenm, perhaps, there are reasons for making redundant structure.

Zabbix Cluster Database ID conflict resolution

The major problem for clustered database is conflict resolution. In general, “an application has to be aware of clustering and must have it’s means of conflicts resolution”.

I have decided to do cluster granularity host-based, there are many reasons for it, and the most important one – is by Zabbix architecture one host must be processed by the same server. This is due to preprocessing, items dependency, and less obvious reasons like proper polling planning to prevent host hammering.

Since different servers will update different hosts, there seems not to be a problem or a source of conflicts or data integrity.

Except logged things like audit, and most importantly events and problems tables.

When I have tried to launch a “cluster” – two instances of a server on a single database, the first problem I saw was that IDs conflict when adding new events and problems.

I guess that to have a database compatibility or maybe for some historical reasons Zabbix doesn’t rely on database ID generation and instead uses its own C function to generate the new ones.

And this is the place where changes are coming. In theory there are several approaches to maintain unique IDs along the cluster:

• Generating some long unique prefix (or UUID) and using it in ID by adding locally-unique counter to it. This approach has benefits of being automatic (no need manual SERVER ID setting), but it doesn’t fit to the existing database structure.

• Stepping: Each SERVER is assigned a sequential ID – SERVER ID(say 0….7). New row ID is generated as lowest exceeding number which is multiple of number of servers then SERVER ID added to that number. ID’s are never will be sequential in this case.They will have  gaps, but we are having 8 bytes for INT now, who cares?. The disadvantage – need to assign IDs manually to each server.

• Ranging: each server is assigned it’s own range of IDS: say,

• server1: 0….1000000,
• server2: 10000001 – 2000000
• and so on.

This might be calculated based on server ID which like in previous case could be sequential 0…7. The disadvantage is that existing methods and some logic of ID generation will not work and more complicated queries to initial ID retrieval will be required.

So, the only real compatible and feasible way is to implement IDs generation for Zabbix clustering is to use stepping. And it’s turned out to be simple, only a few lines of code to fix actually

Zabbix cluster and the “CAP theorem”

Zabbix cluster assumes having some sort of database redundancy. Having clickhouse patch makes things a bit easier – amount of database writes reduces to drastically.As clickhouse by it nature is replicated, there is no problem with history storage.

Consistency:

But for the rest of the data there are some issues as Zabbix keeps state in the database:

• It uses a kind of redundant trigger state storage – triggers states, events and problems
• It updates items reachability status in DB

So running Zabbix on the cluster must to deal with such a data inconsistency as it must be assumed that just before a server went out of service it didn’t updated it’s database, or it could do a partial update:

For example, a trigger state is recorded, but no event or problem generated.

So I did the following: when a new host is assigned to the server, it will reset it’s items states to “Undefined”, so next trigger recalculation will cause new trigger value to be written and initial problem and event records will be generated. This may cause some repeating events and problems, but guaranties that the item is in proper state in the DB after first trigger calculation.

Availability:

This is up to cluster design.Since I am on PostgreSQL right now, I have decided to use BDR. It allows to have own database on each server and promises to have databases in the same “eventually over the time”.

Partitioning tolerance

The point to have own database server is to be tolerant to cluster partitioning. I assume the database is installed on the same machine as the Zabbix server or close by it. So on event of the cluster partitioning server to database connectivity remains alive. Not really sure what will happen to Zabbix if it’s not, it definitely has some means of waiting for the database to become alive. And perhaps till that stops metrics processing as history syncers would wait for the DB to come back to write new trigger states.

Overall

Zabbix fits very well to cluster which works by distributing hosts among servers. All CAP theorem problems aren’t big ussues. I assume even having history being writing to SQL database is still OK for CAP consideartions.

ZABBIX: the "UI meltdown prototype"

Zabbix's UI speed problems have the same root as the database problems - they try to combine two worlds:

 - online situation
 - logs and statistics data from huge OLTP database.

Combination of these two have kind a good informativeness, but at the same time it is slow due to the need of selecting big volumes of data.

So to make things really fast and to have a monitoring which works when everything else fails(and when we need online monitoring most), we need to split it.

The "analytical" and slow part of all the data - like acknowledge statuses, messages, confirmations, alerts data must not be used for online reports.

The proper place for such an information is the database.

Preferably non - OLTP, but BigData one, which is already used for history purposes. But since it's not that much of such a data,it fits ok to OLTP DB for now.

So finally I've got some time to implement the idea.


The Idea points:

Firstly, the server depends even less from the database.

Secondly, online monitoring will be less depended from a  database, only in the server's memory, which have very nice implications, which i write later.

Thirdly, on crisis times (huge network outages), when we need fast online info at most, the database will have a hard times due to large amount of events and problem update records going into the database, so we'd rather have out-of-DB fast way to know what is the current monitoring state, while we have time later to analyze events and problems.

The coding

I have created one more queue in configuration cache, named "problems". Queue updated each time on event when Zabbix's export problems feature is invoked. On recovery events problems are deleted from the queue.
 
Since i needed id indexing and no sorting at all i decided to go with zabbix's hashmap librairy and used hashing based on the eventid.

In the trapper code i did one more message processing:  
{"request":"problems.get"}

Processing is rather simple - iterating over the problems hashset and creating huge JSON response.

I did locking there, since events are registered from different threads (i assume it's the preproces manger which calculates triggers) and  export seems to be happening somewhere else (export thread?).

And it was the easiest part.

Fixing the UI was quite a deal.

Zabbix team used MVC ideology, so there are three scripts to fix and different types of API colls are spread among all three of them.

And there are separate classes to learn for Server and html construction.

Actually, it is very nice from code and structure point of view. It's a "heaven for perfectionist" programmer as it should be, but i wasn't quite ready to fit it all into my brain at once. Whatever, it was fun anyway.

The result: i was able to get stable UI response time to less then 100msec, having about 50-70 msec of response wait time. I've also realized that it was just enough to remove all that nice pop-up boxes with problem descriptions, acknowledge status to have OK response time on  DB load during huge number of hosts outage retrieving data from DB  without trapper. It's more close to 2000msec but still acceptable. So a kind of "easy" fix is also possible.



The problems: sometimes PHP cannot parse output from the zabbix server. Dumps show that PHP gets all the data, i can see buffer in PHP logs, but parser doesn't parse it. Due to rare nature of the thing i couldn't find its roots yet.

Another small disadvantage is the problem start time - as i keep the time of the event, usually it reflects the time when zabbix_server was last restarted, which happens a lot on devel times but rare on production.



The big advantage- it's fast. No, it just _really_ fast! 

It feels just like opening a static page. And it is really doesn't need php-fpm resources. Which means we can give up separate frontend machines, it's not that critical but just good.

So, this is full win in terms of prototype and usability, but it's a bit non production right now, because fixes have broken the "problems" panel and the changes are done by fixing existing widgets, instead they should be added as a separate ones.


Some tests:

I've tested two important on-line monitoring widgets - problems and problems by severity under "average collapse" situation (10k hosts are inaccessible).
Numbers are time in msec till widget has been rendered. For trapper widgets - no difference for ok and 10k fail situation.

On the full 50k+ collapse both reports couldn't fit into 300 seconds.

I've decided not to waste time to allow nginx wait for fpm forever just to figure it maybe will render in, say, 12 minutes.

I'd say that 2 seconds is acceptable, 10 is bad, more than 10 - no chance for production.


And there is one other great thing about this - now we can get online zabbix state without db.

If and when zabbix OLTP mass updates like events and problems will go to BigData DB, then it will be very-very close to becoming a real clustered server:

The idea: Two (or more) servers could use trapper interface to make sure they both functional and split hosts load between them: odd hostsids to one server, even ids to the other. Both serves will have full items database in both db and memory, but on code which distributes tasks to poller one condition should be added - not use the other's servers hosts tasks until it is alive. So when it dies a hosts will be activated and polled

Sure, there are complications: LLD items, active agents, zabbix senders, but i think it's solvable.

ZABBIX: the story

Long long ago, perhaps two months for now i have started one week project that seems to be finished only by this fall....

So, ZABBIX.

At the job we use it for monitoring all the nice infrastructure we have. And It happened so that the monitoring service was growing but didn't have a true owner  last four years. So i decided to put my hands on it.

There was several reasons

• technical problems (hangs, slow speed, short history period)
• it was quite outdated

but the primary ones where:

• it's a fun,
• it's a new knowledge,
• because "i can"
• it's a real task that's make a lot of value for business

So, before i started to do something, there was a time for thinking, circling around, drawing schemes, learning new versions of ZABBIX and getting access to the production monitoring servers. That lead me to understanding idea that monitoring consist of two major parts - real-time and analytical.

At the end of May, 2018, on the internal hackaton event i completed the clickhouse history module. I would name this the most important change and "must have" to everyone who doesn't use ElasticsSearch. But even those  who does please read a comparison. Overall, clickhouse works very good as a historical storage. It is quite easy to setup and it makes a big deal. 

To make ZABBIX support the Clickhouse storage you have to be ready to compile ZABBIX from sources with patches. Also clickhouse must be setup along with initial data scheme. Look here how to do it.

But the story didn't finish then. After completing the history storage part some other problems come to my view. It was the number of servers we used for monitoring. We had twenty one!, fifteen of which where proxies to keep up with the load.

A long and curvy road led me to optimize that significantly. I was under impression that most problems came from constant forking of fping process. After some research and  efforts to reduce number of forks per second<> (fps :) ), i did the nmap solution  along with "time distributed pinging" idea. Overall that allowed to make 30-40 times less forks while doing checks three times faster and save about two times of CPU ticks on accessibility checks. .

Then it was SNMP querying. Comparing to accessibility checks we have ten times more items to poll. There are almost 3.5 millions of items now. And polling all in reasonable time frame couldn't be done with only one server even with 1000 threads.

Before I did several  projects with mass SNMP polling on Perl and Python, tried AsyncIO or plain threading, so i've already knew that asynchronous polling would do the thing. Description of the asynchronous polling idea  is in this  article. 

To implement mass polling there was many things to be done in the ZABBIX server beforehand. ZABBIX is somewhat ready for batch items processing. But there are many places in the code where processing is done on one-by-one basis quite ineffectivly. I named this part of job "BULKiing"

When processing batches of items was completed it is also became apparent  that there are two major architecture bottlenecks. One is preprocessor manager with its "queuing problem" and the fact that all the items being processed by single thread, the other problem is global locking of configuration. Both problems lead to a situation when systems with  slow CPU cores will perfoms not really well running ZABBIX under high load. Such a systems show low processing NVPS numbers with while most CPUs being 80-90% in idle state waiting for internal locks.

Both bottlenecks are limited by single CPU core speed. To somewhat avoid preprocessor manager stucking during high load i changed the way it processes  data so it could prioritize  tasks and avoid queuing. Now it can decide what is better to do - getting new results from pollers or sending results to workers. Details are here.

To solve "global locking" i've attempted to enlarge batches of items to poll from 8k to 32k which gave about 10kNVPS performance increase and allowed to marginaly exceed initial goal of 50kNVPS processing. But after we finally put the modified version of the ZABBIX to the production the problem with slow core and global locking appeared again. 

Even it could be solved with faster CPU, i decided to solve this either. This would allow to fully utilize older server hardware with slower CPUs but large number of cores. Detais are here.

And then to resolve another "bottleneck" problem, i've changed architecture to process data by more then one process manager. Averall together with all previous fixes it allowed to raise total processing capacity from 5k to 88kNVPS on the same hardware additionally eliminating need of proxies we put for load share.

The same code could show 118kNVPS on the faster-core test machine.  

Lastly i will be doing some UI-speed related work due to UI not functional under stress situation and even in normal conditions it takes 5-10 seconds to refresh panels.

And to make things complete, there are several short notes which are more about mood and fun and strange situations, architecture questions, lack of knowledge,  coincidences, links to sources, some work records, and so on.

nb: All the numbers are given for our installation of 47k hosts/4M items. I assume many of the problems are not likely to appear on installations with less objects.

nb2: The primary reason i write this is the possibility to catch a glimpse of engineering mood and fun. I accidentally found my 6 y/o records from the time we where launching new BRAS type. Reading that was kind a joy of reliving that time of being consumed with engineering, ideas, solutions, sleepless nights. So i am writing now that I'll have something to read in 2024.

ZABBIX: DZSA (direct zabbix server access)

The essential UI problem is that it uses API.

The problem is in efficiency, especially on downtime times. Especially when DB is slooow under 100k+ problems, events, triggers updates. (Remember the crash?). Most important widgets looks like this most of the time:

Solution:
Have a list of active problems on the Zabbix server side, refresh it according to problems happening or closing on the "go".

Fetch it via trapper interface.
Build  "current status" widgets out of them, avoid API as much as possible.

Side effect:
Have a list of "internal problems" for monitoring system administrator to see what's working wrong in terms of monitoring, use the same idea to build such an interface.

The progress

server->trapper->php->browser "hello world" prototype seems to be working:

Lets see how it works all together completed a bit later.

ZABBIX: nmap, fping, collapse reasons

Today i finally had time to figure what's happened on the zabbix crash two days ago.

The reason so much hosts stopped to be accessible was the nmap fail. I've fixed it by switching to fping, but i had to figure out, why nmap has stopped working.

During the crash we've tried to fix networking accessibility problem and  someone has left a nat rule to map ICMP traffic (most networks) to the proper src address.

And this is important: i dont' know the exact reason, but it seems that nat rule set on POSTROUTING stage HAS been applying to traffic, and returning traffic for some reason hasn't been able to be received by the same socket. I am not sure if it went directly to the socket without being deNatted or it couldn't get the socket after being deNATted by iptables. Doesn't really matter.

Fping was functional under same conditions due to it was already sending traffic from the correct address.

For nmap situation is a bit more complicated: when the -S option is used, it doesn't actually sends any traffic. To be precise, no traffic is leaving the system with the desired destination. Only setting outgoing interface helps (perhaps, one more reason to learn RAW sockets nature).

So to utilize source address with NMAP functionality there must be outgoing interface to be known by zabbix server or outgoing interface has to be set in the configuration. Latter seems to be simpler.

And one more thing to consider: actually, nmap parameters where wrong, so alongside with icmp it was sending packets to 443 and 80 port for host discovery. This is not right as may produce harm on slow devices, so I've fixed the nmap options.

ZABBIX: v4 in productions and first production collapse

We've got the v4 working for production now.




While we where trying to fix the nmap source address problem, we accidentally disrupted network connectivity, which lead to the mass system collapse:

The core reason of collapse is slow mysql update speed. This, in turn have roots from ZABBIX nature to log everything to the MYSQL database.

And third, by architecture, Zabbix server and frontend are exchanging data through Mysql.

What happens is:
Net collapse-> lots of updates (problems, events, etc) -> db collapse -> slow frontend -> more db load due to page refreshes.

Some thoughts:

1.Do exchanging of on-line data (existing problems) via some in-memory method.
2. Reduce as much as possible DB load and logging of events, and (probably) send them to clickhouse.

I have also found that there is a huge trends_uint table. I thought that it's housekeeper's job to update it. Since there is HouseKeepingFrequency=0 in the config file i haven't check it before. So i did "update items set trends=0" to make sure.

But lets' retrun to collapsing on high load.
It's both 1) and 2) needs to be fixed. But also something have to be done about frontend widgets PHP part wich seems to be nonfunctional under high number of innacessible hosts.

ZABBIX: upgrade to v4

So far i am very glad to have started the upgrade. There are many changes in the code.

The most important now - they redid locking. However  it's kind a diffrent from i did, but, perhaps more right.

Instead of splitting queues to pieces, i saw that now there are different locks for read and write. That sounds to be promising as now it will be possible to use reading in parallel. Perhaps its good, but testing didn't show any performance increase yet.

And i don't remember this before:

It's from DCconfig_poller_get_items - function to retrieve items from queues to check. I think it's just not right.

It's much better to make async requesting in every type of check. Do some reasonable timeouts and just _forget_ about priorities, different timing and unreachable pollers.

I tried it for snmp and saw no negative impact.

Whatever number hosts are not reachable the thread will wait only one timeout of 2 seconds either for one or eight thousand hosts.

Some disadvantage of async pollers is that in theory they will  work a bit longer for pack batch of hosts then they would do for a single item. Each pack of items might be processed from 2 to 8 seconds, so it's not good for polling data more frequent then 10 seconds.

But according to my tests end expereince, its not worse then existing system, under comparable NVPS it will be very fast, but don't expect 1-2 second polling delays from the system processing 20-30kNVPS

ZABBIX upgrade to v4, some pure load test

As a final part of work i want to upgrade all to v4 line to be able to support the changes as long as possible with minimum efforts. (And to work as early adopter and bug fixer for Zabbix)

I have also decided to do some pure untouched virgin Zabbix load tests.

I took clean v4.0.0 alpha 9, applied clickhouse changes.

By the way i saw some minor elastics changes, but what is more important - they seem to be implemented file offloading of events, history, trends, perhaps something else.

That is valuable. Perhaps if found this before, the clickhouse offloading would be implemented outside if the main server process as a file parser. On the other hand, current implementation is in process, efficient and also frontend-compatible.

The reason to do load test is a to find that maybe Zabbix will be able to do out of the box what we want it to do.

This time i've tried to fetch as much SNMP traffic as possible:

pic: (1) system is 44% loaded, but (2) collecting only 21-22k new values per 5 seconds, so it's 5-6k NVPS. It's spends 10-12 seconds (4) in waiting for configuration locks, under normal conditions poller processes spends 4-5 seconds per each 50-60 values. Totaly, there are 3500(!) (4) poller threads running. And they also exhausted most of the memory.

I couldn't make it run more then 4k threads out of the box, but assuming linear CPU growth it will be able to grow to 8k after some sysctl fixes and gather 12k NVPS in the best case.

Then i've decided to test again pure accessibility check speed. As i mentioned before, i could acheive 28k NVPS on the same machine of pure accesibily 1-packet check on fixed icmp module which utilizes nmap instead of fping.

Ok, pure fping. System is quite loaded. Load average is 466 :), no idle time, 10-11kNVPS performance.

So, Zabbix cannot do alone out of the box what we need and there is a reason to apply fping and snmp patches, since they really make it possible to collect 15 times more data on the same hardware.

Wow, Thats Funny! 100k NVPS exceeded!

Subject is inspired by on of the most popular TED YouTube videos.


I was recently doing load testing of the new code altogether and found a funny thing:

After certain point lowering delay for items leads to dramatic degradation of preprocessor manager.


I've  added statistics logging to worker threads and found out that they are free most of the time, so this time it looks like it's the preprocessor manager problems.

After reducing items delay to something less then 30 seconds i saw that preprocessor managers drop their processing speeds from around 80 000 items to 500-800 items per 5 seconds.

 

pic: semop 4 shown, but real problem is the semop 1

Perhaps, it's still locking. There aren't that many semop operations, according to strace, but visually it's the operation where the process freezes most. I couldn't find a simple way to calculate wait time via strace or something similar, so have to rely on my eyes. (update: RTFM, -w  -c options combined  produces strace statistics based on real time)

First i've found prerproc_sync_configuration was guilty, but this was wrong path.

After an hour of research i've found another bottleneck. And this one is also global architectural - it the history cache locks.

Situation is almost the same as it was with configuration lock. There is a globally locked access to history storage which leads to "single CPU problem"

Having LOTS of results coming from pollers, threads become limited by a single core speed to put data to the cache and to read it from.

Solution? Well, it nothing new - lets split it to 4 independent caches which is locked separately and access each one by it's sets of threads.

Most of dbcache.c has to be fixed as it's not only the cache that has to be split into 4 parts, but there are lot's of static vars declared.

Not sure if this is something needed for production but a might be a good prototyping experiment.

It is also a question hanging - why performance raises when i split preprocessing manager to 4 workers on Xeon5645 CPUs? (upd: perhaps the answer is that preprocessor manager has lots of other processing work to do which being split by differnt CPUs give some combined profit)

Whatever.

Just to make sure the job of maximizing NVPS is completed,  i fixed dbcache.c and HistoryCache locking to support 4 hsitory caches to see if this is _the last_ bottleneck.

I did what i assume safe locking - rare and non critical operations are still lock the entire cache operations, but most critical ones - only locks it's part of the cache.

Data flush from preprocessor manager to cache, where speed is mostly needed, are done by locking one of four sub caches, and such an additions can go in parallel. Then i also had to fix lots of static vars in dbcache.c as they seem to be used in parralel.

But i have to admit i've failed on this experiment. Spent 3 day's free time, but still was getting very unreliable daemon which keep crashing after 1-2 minutes.

So at the moment i am giving up on this, perhaps later i may revert to this.

pic: four preprocessor managers, doing 118k NVPS combined (note - idle time calculated wrong)



And there is one more reason.

I returned to test machine with E3-1280. And wow!, Zabbix could poll, process and save 118kNVPS (pic) with fixed queueing and 4 preprocessor managers. I would say its 105-110k stable while still having 18-20%idle. And It's kind a lot. My initial goal was to reach 50k. Considering we need 10k on one instance and 2k on the other it's enough capacity to grow for a while.

ZABBIX: multiple preprocessing managers

I decided to do that due to one thread of Xeon5645 couldn't handle all the preprocessing job, and also this might be required for faster CPU's but for higher loads.

First, some thoughts i had ahead:

• 1.there might be a problem if two dependent items go to different managers so they will be waiting one another (deadlock)
• 2. It must be some kind of workers distribution.

My first idea was to split requests and responses to threads, but then i decided it will be much easier to solve 1 and 2.

Since there was already host-to-queue persistence in queuing version, so i did thread-to-preprocmanager persistence the same way (procnum%4). So all items from each host will get to the same preprocessing manager.

Workers are split by preprocessing managers in the same way, by hashing thair procnum.


So, after fixing the worker log handling problem, preprocessors could handle much more requests.I saw stable value around 110k per 5 seconds per each manager, giving altogether 88k NVPS. I thought of making a 100k, but decided not to waste time on this old hardware, but test something more up to date later.

By the way, measuring prprocessor manager statistics is the way how i prefer to measure NVPS. I assume server collects a bit more from the network. It does requests to inaccessible hosts, which return no data and don't appear at preprocessing manager.

So thats it. Splitting preproc manager is essential if there is a need to achieve more then 50k NVPS stable on modern CPU's. This is what actually was limiting the test server with Xeon E3-1280.

ZABBIX: the second server

Tried to launch ZABBIX with all the fixes on the role of servers monitoring, started to segv.This was a bug in error logging.

There is another (big?)
problem there:

 

It looks like poller needs to be further altered to support asynchronous checks for agents.Or there some kind of locking occurs.

Strange thing that simple checks are slow too. Or might be global thing like DNS or something

This installation is rather small: 

And the server is sure enough idle 97%.

upd: nothing interesting, new server was not permitted by ip filters to access hosts via SNMP, and for agents there was also no permittion to it to request data. 

ZABBIX: what does a cat when it have a free time?

subj. ZABBIX does something quite opposite.

I had successfully changed the code to have as many preprocessor managers as i want.

But whatever i did, stuck at 60k nvps.
My initial thought was it's preprocessor manager slows things down. So i've rewrite it to have 4 managers.

But then i saw that running 4 managers spreads CPU load between them, but gives no NVPS increase.

Whatever i thought, i couldn't find a reason to such a behavior. But today it struck me - it's worker threads and workers are waiting for them.

So, i straced worker thread.

.. semop....
Ok, is it GLOBAL config LOCK again?

NO, it's mutex 0, the mutex dedicated for logs.
But i don't see anything from workers in logs and it is loglevel is 3.

And guess what?
Let me show you the main loop of worker code:

 

I commented it out. It does log rotation. With locking. Inside loop which runs possibly in 100 threads and which task is to process an item as fast as possible.

So all the worker threads was keep trying to put LOG lock and to rotate the log file if it's bigger then config values. I guess it's a developers mistake. 

So, without log handling workers could do a bit faster and server achieved approximately 85-90kNVPS, while having only 25% idle. 

So, perhaps, that's the limit for double  E5645 system with 12 physical cores (and total 24 cores if we count hyper threaded ones).

During all this tests i've also found preprocessor manager priority flaw - it was more likely that it will gather queues before the buffer is full. And when the buffer is full it would be working OK, but it would be slowing down pollers.

Now it's just fine. Queue is close to zero. Each one of managers writing about 110k items each 5 seconds. So totally it is 110k*4threads/5seconds=88kNVPS.

ZABBIX: crashtesting, records #2

....failed very fast, on lots of process the system starts to swap

fixed the buffer, the system does fine with 20 snmp threads with, about 25knvps

have significant queuening with 43k required nvps, try to make 30 threads


didn't help, changed delay from 90 to 120, zabbix somehow stopped to work normallu, restarted.

doesn't work either, reduced numbre of pollers to 20
queueing, 33nvps.....
raised pollers back to 40....

OK, 30NVPS ONLY SNMP - NO PROBLEM
changed back delay to snmp to 90s.... lets look
it still queueng but just above 10sec, at planned load of 43knvps

reduced delay of snmp to 100 sec,pinger still off, let' see
ok, the system is stable to complete zero queues at 33kNVPS of clean SNMP load.

Now, let's test the FPING 100 threads:
the old version could handle the load without big broblems with up to 1sec (1sec had quieuieng in the first column)

NMAP: 100 threads works worse then FPING
high system load,  32% idle

Reduced NMAP threads to 20:
It does just like 100 threads FPING but system idle is 62%

Reduced NMAP threads to 10: it's worse, 20 seems more optimal, it's 82% idle

Now try fping in 20 threads:

The normal SNMP is:
0:00.26 /usr/local/sbin/zabbix_server: poller #1 [got 249 values in 5.005509 sec, getting values]                                
 7942 zabbix    20   0 13,670g 271628 213924 S   0,3  1,7   0:00.26 /usr/local/sbin/zabbix_server: poller #4 [got 175 values in 4.851671 sec, getting values]                                
 7943 zabbix    20   0 13,670g 265808 208076 S   0,3  1,6   0:00.25 /usr/local/sbin/zabbix_server: poller #5 [got 142 values in 4.914348 sec, getting values]                                
 7946 zabbix    20   0 13,670g 268228 210508 S   0,3  1,6   0:00.25 /usr/local/sbin/zabbix_server: poller #8 [got 175 values in 5.141629 sec, getting values]                                
 7947 zabbix    20   0 13,670g 274664 216944 S   0,3  1,7   0:00.26 /usr/local/sbin/zabbix_server: poller #9 [got 158 values in 5.044240 sec, getting values]                                
 7949 zabbix    20   0 13,670g 264984 207320 S   0,3  1,6   0:00.24 /usr/local/sbin/zabbix_server: poller #11 [got 184 values in 5.038127 sec, getting values]                               
 7951 zabbix    20   0 13,670g 273344 215680 S   0,3  1,7   0:00.27 /usr/local/sbin/zabbix_server: poller #13 [got 138 values in 4.924217 sec, getting values]                               
 7953 zabbix    20   0 13,670g 287580 229860 S   0,3  1,8   0:00.27 /usr/local/sbin/zabbix_server: poller #15 [got 172 values in 5.133177 sec, getting values]                               
 7955 zabbix    20   0 13,670g 258828 201032 S   0,3  1,6   0:00.24 /usr/local/sbin/zabbix_server: poller #17 [got 142 values in 4.982964 sec, getting values]            

140-150 per thred per 5 seconds
tried to make 300 threads



fping synthetic results:
 fixed fping
real    0m42,646s
user    0m0,236s
sys    0m0,336s


total CPU usage: 0,572s

fping normal

real    1m33,304s
user    0m0,104s
sys    0m0,276s
total CPU usage: 0,380s

nmap
real    0m3,252s
user    0m0,220s
sys    0m0,076s
total CPU usage: 0,296s

ZABBIX: crashing, getting syssegvs and so on -dairy records

And still i see the server is failing (segv sometimes).
It's not nearly as stable as the production one but tipycally these problems are quite easy to catch

The problem now is that server restart is taking more and more time (about 2-3 minutes)



The preprocess manager problem



The questions:

"something impossible happen logs"

why snmp is slow to be enabled? - figure why unreachible threads are waiting for nothing



huge load after some time!!!! net_snmp_close()???

in the night had huge number of must use snmp_select_socket2 (or so messages) - seems that under certain conditions i leave opened SNMP sessions.
Too many sessions cause high load and what seems to be select() call is not working, so essentially  all the gathering process breaks
TO DO: add some profiling, look where i may leave sessions unlclosed, consider removing (select) code and replacing by stupid simple wait call.


morning: huge snmp queue, seems that threads work normaly, i haven't found any abnormalies, just 10 them is not enought. (actually, they work much longer then at start, their run time raized from 6-9 seconds to 14-25, but that might be due high system load)

Raised snmp pollers number to 25, starting at 08:08
8 58 flight normal queue 0

seems something strange has happen at 11 24-11 34, apparently serever has crashed at that time, there is a long queue in SNMP in the browser left
11 41 started again ... actually the server is alive, but is't swapping and there are 16M records in the queue of the preprocess manager.

Myabe that's just too much hosts???
Maybe that's too big load from fpings
Now i disabled about 20k hosts, reduced number of threads to 20 for both pingrer and poller, restarted  the server, let's see what's happen


Having only half of the load (and also number of DB Syncers increased to 40) system is seems to be doing fine for at least 1 hour for now. And this is the time when i finaly see some SNMP data gathered in the graphs.

I will look at it for another 2 hours then will start gradually adding new devices.
UPD so far evth is fine, no queueing, system 90% idle


Another problem to consider - devices marked non - snmp acessible stay too long in such a state and i've never saw any processing done by unreachible pollers.
I exect i've broken some fetching logic in DC functions and that makes it such a long to mark devices as SNMP accessible again


So my theory is: 1.  I haven't enought data syncer processes enabled before so they where constantly busy in filling the value cache (ClickHouse not really fast for single host queries). Acutally i would suggested to fill the cache on the start by some bulky way - this would mean one query with millions of data. Unfortunately, i am not sure if this is what really zabbix and it's triggers need.


 The biggest thing that bothering me right now is occasional preprocessing manager queueng. I am almost sure that the queuing depends on two factors: system load and number of working threads (actually it might be one of them, and this two factors are closely linked). There one big optimization that could be done on the SNMP poller - is to enable bulk queriing, this will save some ticks of CPU and network PPS, might be easier for network devices, but it will not faster things really as anyway each SNMP thread will be waiting for inacessible devices timouts which more longer than making 100 queries to accessible device, and they happen in parralel in the asyncronious model.

Another big change and improvement that might be made is switching from fping to nmap for hosts checking. Question one - is how valuable the packet loss rates for understanding device reachibility. I have a thought that it might be more valuable to make 10 pings each 3 seconds then make them all at once each 30 seconds. In such a way of checking the packet loss rate will be done in a trigger by calculating average on base of last 5-10 checks. To speed up reaction of a device has become non accessible, last 2-3 checks must be considered.


ok, morning, 3 17. I see that the installation has crashed in preprocessing manager caught syssegv.

All pingers has been off (i've set up StartPingers=0)

I really believe that it's preprocessing manager that has to be blamed, in particular,  queueing of items. So, i'd like to fixid to one of the following
- find an answer - why is queuing happening at all???
- do not allow queueng
- after certain amount of queueing throttle the processes to wait till preprocessing manager be free (if it's load problem, which seems not)

I have nice thing - i know that running zabbix in debug mode almost immediately causes the queueuing. So i need to trace down all the decision process to understand, why the hell does it queues the messages


I see that most of it's time it's spend processing IPC requests of type 2 (probably, that the result of the poller's work). added extra flush and history flush logging to see if it's happening at all. On the next step will add _flush call result

the Github and sources, naming

The sources are in the two projects at github
Cleaned and checked sources and patch files are at
https://github.com/miklert/zabbix
At least separate clickhouse patch will be there soon.

The very unstable development version is at
https://github.com/miklert/xe-rabbix

Now it's based on 3.4.9 code.

It's a few words about compiling zabbix - you'll need MariaDB-shared library, not just client and server libriares and development files. I've spent quite a time finding that out.

For clickhouse compile with curl libriary

Regarding the name. We had to call it somehow internally, and every monitoring system must have something common in name with ZABBIX.

So we started to call it xerabbix [herabbiks] first, which appeared from Russian soft-sounding word 'dick' and ZABBIX, but it was too rude, so after a month of development it's transformed to xe-rabbix and now pronounced as [iks-e-rabbiks], leaving  rudeness in the past.

Nowdays we interpret the name as eXtra Edition of zabbix.

ZABBIX: Configuration cache, DC_poller_get_items, DCpoller_requeue_items, queues

A test machine was able to achieve 51kNVPS steady..

It more then enough. Test version could run for more then a week at 16kNVPS stable, no memory leaks. So we have  planned to put fixed ZABBIX code into production.

The old installation was ZABBIX 2.1 version. The migrating job included lots of other works like changing OS, fixing automation. Long night work.

Just after the server was upgraded to new OS, we started the database and the server and .... OOOPS it could only do 5kNVPS. The CPU was 90% idle at that time. We have seen this already, right? To finish the job the new server left work for  production on test machine.
 
Ok, lets figure what's going on the new server.
This is statistics that strace shows for a poller thread.

It clealry shows that threads  waiting for it's turn to get a lock .


Now a few ideas about locking.

ZABBIX holds in memory huge configuration structure, which it holds in several tables (perhaps it's lists or trees).

Apart from that there are 5 queues of tasks by polling type (ICMP, POLLER, PROXY, so on).
In each queue there are structures ordered and (perhaps) hashed by next check time. The structures references items.

So, each time something gets access to the configuration, it sets global lock via mutex and semop, reads/updates configuration and then unlocks.

 The global lock domain picture:

After analyzing the dbconfig.c and dbcahce.c code,  i decided, that for some operations global locks might be avoided.  To be precise - all operations from poller and pinger are safe to go in parallel as soon as they use and lock their own queues:

I've decided to split each queue type into 4 pieces. To keep host-per thread persistence items are distributed by queues by (hostid%4) hash.
To maintain polling threads persistence, threads are bounded to queues by their hash (procnum%4).

It should be at least 8 threads of a type (but better 16 threads or more), so at any time at least one thread could request data from DC cache, while others might be doing polling job.

So the result?

But before i'll tell about one funny twist in the situation:

After fixing the queues and making them work i see that the server performance is still not better.

Strace shown that poller still spend most time  semop calls waiting for global lock (mutex ZBX_MUTEX_CONFIG, 4).


2512:20180723:121820.512 In DCget_user_macro() macro:'{$SNMP_COMMUNITY}'
zabbix_server [2512]: zbx_mutex_lock: global config lock attepmt!
zabbix_server [2512]: zbx_mutex_lock: global config unlocked

After doing some profiling i found that  problems was in the macro nobody really needs, it's macro for community, and we only have one community anyway!

And then i even thought that maybe ALL the problems with locking  where due to this macro and one week of coding was just due to my lack of  knowledge of ZABBIX.

So, i reverted to no_queue ZABBIX version. It have shown that performance is much better without the macro, but only 2 times.

Still no good, so were going the right way with queues.

As we are in the development, lets do fast fix:

update items set snmp_comuntiy='isread' where snmp_community like '{$SNMP_COMMUNITY}'

(perhaps for production it's better to use API or UI to fix templates).


The following pic best describes the first launch of ZABBIX with queues:

On the same server which was doing 5k NVPS before,ZABIX shown 60k NVPS steady.

And finally for the first time ever i could load machine to 50% with ZABBIX.

Why no more? Because then next limiting thing comes to play - preprocess manager.

ZABBIX: preprocessor manager queue buffer

This happened first time right after I've launched asynchronous SNMP polling, which immediately raised NVPS from 2-3k to 7-10k

First I've noticed there is a significant data delay from polling to seeing it in the ClickHouse database. Doing tcpdumps and selects from Clickhouse revealed that data stucks inside zabbix server daemon. Since tcpdump shown that from network traffic data was gathered in time, pollers where not to blame.

Looking to the top -c output i saw that preprocessor manger has queue parameter information set by setproctitle which where constantly growing.

So my first job to find why it was growing. As this was development machine i had logging set to debug all the time.

Logs didn't show me where the process manager was spending all its time.
After I added some extra logs and fixed log.c to increase logging time precision to 1/10000s.
I've found nothing, but the strange fact that all the operations takes 190-222microseconds. And that was interesting because it meant something long happens each time i write information to logs. Right, it's logging itself.

After i switched off logging queuing disappeared, so i thought I've solved the problem.

However a week a so later when systems started to reach 20k NVPS i saw that preprocessor manager starts queuing again.

Queuing is a good thing as it provides a buffer for overload times. But in my case it was happening endlessly, queue was growing all the time, until ZABBIX couldn't allocate more data.

I had to start investigation again. And this time it was tricky since i couldn't use logging of every operation as logging introduced much more problems and  overhead then the problem i was looking for.

Finally, from reading old logs, looking at the source, the problem was found:

Process manager has a socket where it accept two types of data. First, there are request from pollers to process newly collected items. Requests then placed to queue. And then preprocess manager distributes tasks for preprocessing to worker threads.
The second type of message is worker responses with the data that has been preprocessed.

With big batches of collected items (at that time they where 4k size, later they increased to 32k) preprocessor manager had no choice but to constantly queue items: when a big chunk of data arrived to the IPC socket, it read all the data and then sent tasks to process items to worker thread. As it is much more items in a batch then worker threads, it had to queue most part of the batch. Then it has to read from socket again since to release worker thread to do it's next item, it has to submit the result first.

(Note, fixed later: about month after i accidentally realized that message sizes where limited to 256 items and the only factor for queuing was processing speed)

I've altered the code to open the second socket so preprocessor manager could decide which data to read - values from pollers or results from workers.

This allowed to survive load peak times when processing of new data was going slower then pollers could collect it.  And the bigger queue was, the more processing time needed (or it might be that swapping started at the test system, i had some memory constraints on the test system).

The fix works the following way: preprocess manager works the usual way until it gets 1M records in the queue. Then it stops reading requests from pollers and only sends data to workers. Until there is less then 1M in the queue.

Pollers have to wait untill their request processed, so on peak times when queue is full, polling for new data slows. This is much better then eventually crashing with data collected. I saw queue reaching 20M items.

Preprocessor manager is the only bottleneck left for now, so i consider later to either duplicate it or split functionality to two threads.

One thread will process requests and the other results.

Making two equally functional thread seems to be more difficult as we may fall in the situation when depended items are processed on different threads, making them wait endlessly for the other item to arrive. Or they have to share common storage and use mutexes to arbiter access to the storage safely.