Hi all, I wanted to chime in briefly to say that I appreciate everyone sharing your experiences with high CPU issues, as it does help the developers & I to hear what everyone is encountering under heavier load and/or bot activity. The more that institutions can share your experiences, the more likely we can begin to narrow down the problem(s) and build better documentation/guidelines for everyone.
A few things that are clear is that Server Side Rendering (SSR) from Angular **does seem to be more CPU heavy than we anticipated**. This is why the basic "SSR caching <https://wiki.lyrasis.org/display/DSDOC7x/Performance+Tuning+DSpace#PerformanceTuningDSpace-Turnon(orincrease)cachingofServer-SideRenderedpages>" was added in the first place. However, what's also starting to become clear is that the *basic SSR caching* may not be enough. (In all honesty, we knew it would help in some scenarios but possibly not *all* scenarios.) I can verify though that the existing basic SSR caching is *per instance*. So, when using "cluster mode" (and running several instances at once), there is no way to currently share that cache across instances (as the cache is literally just stored in the memory of each instance). This means it has a more limited impact than we initially hoped. *This may mean we need to begin looking at some more advanced caching options for Angular SSR*. To be clear though, this SSR performance/caching issue shouldn't be specific to DSpace 7, as we are just using the SSR tools from Angular.io. So, it's possible that tools may already exist out there from other sites/applications that use Angular SSR. In the meantime, I would ask that sites which have this working well consider also sharing your experiences of how you "stabilized" your high CPU. I know there are sites out there who've done this (as there are a growing number of sites running DSpace 7 in production). It'd just be helpful, for me (and others), if we can learn from each other in order to create better documentation & best practices for DSpace 7. (All DSpace documentation & best practices have always been a collaborative/community effort because we don't have a central development team.) Tim On Wednesday, March 6, 2024 at 10:38:43 AM UTC-6 uOttawa Library wrote: > I would like to understand how memory is used by the node instances. > > There are comments in the example frontend configuration file that > mentions the following: > > # Maximum number of pages to cache for known bots. Set to zero (0) > to disable server side caching for bots. > # Default is 1000, which means the 1000 most recently accessed > public pages will be cached. > # As all pages are cached in server memory, increasing this value > will increase memory needs. > # Individual cached pages are usually small (<100KB), so max=1000 > should only require ~100MB of memory. > > We have both bot cache and anonymous cache set to max: 1000. This would > mean a total of ~200MB cache (per instance?). We allocate 1.5GB to > instances (max_memory_restart), so cache wouldn't be the main cause of the > high memory usage. We have set max_old_space_size=1024 since the original > post above, and this seems to make the instances stay alive longer > (instance restart every ~90 min due to exceeding the 1.5GB memory). > > It isn't clear is if cache is shared amongst instances (to avoid having to > render the same frequently accessed content in every instance), but in any > case, it wouldn't be the main source of memory use according to the > comments. > > François > > On Tuesday, March 5, 2024 at 5:13:09 p.m. UTC-5 Edmund Balnaves wrote: > >> We are running DSpace 7 instances in a multi-tennanted environment in a >> reasonably stable way. >> >> Our experience is that lots of memory is needed and we do see lockups in >> cluster instances periodically. Even low levels of bot activity can >> stress the system and performance of DSpace7 is pretty under-whelming but >> we have managed to maintain stable instances. Trimming your caching >> would be wise to keep memory within reasonable bounds. We have written >> shell scripts to monitor and restart instances that look to be locked up, >> and have put some memory limits for auto-restart per your approach. >> >> >> Edmund Balnaves >> Prosentient Systems >> >> >> On Tuesday, March 5, 2024 at 10:22:23 PM UTC+11 Majo wrote: >> >>> Hello Carolyn Sullivan. >>> >>> I would like to offer a few points I noticed. I was responsible for >>> deploying >>> one instance of DSpace and I am quite familiar with problems you >>> described. >>> However I am no expert, so take all of the following with a grain of salt >>> (and perhaps a bit of hope that someone more experienced will also >>> reply). >>> >>> First of all, the resources you have available are by far not enough. >>> The instance we deployed was small and we are using *15 CPUs and 30GB >>> of RAM*. >>> Initially we had about your specs and I couldn't make it work reliably, >>> no matter what I did. >>> (Limiting bots helped a great deal, but that is certainly not ideal and >>> the performance was still terrible). >>> >>> Secondly, the caching. By trial and error, I arrived to 20 sites for >>> bots and 100 sites for anonymous users. When it was significantly >>> higher, each core used a lot >>> of memory and therefore kept restarting, losing all benefit of cache. I >>> suggest you try to observe >>> how much memory your individual cores consume and if it is too much, >>> decrease the cache. >>> My conclusion is, that you do not need exceeding amounts of caching, >>> because yes, >>> it will consume too much memory and if it causes swapping, it will only >>> slow everything >>> down. I was slightly surprised with the settings that work for us, but >>> they are effective, >>> when used together with more CPU and RAM resources. >>> >>> I consider the points above the most important, but what is also >>> peculiar is your setting >>> max_memory_restart = 1500M. Why have 4096MB of memory for each core, if >>> you restart >>> it at 1500M? We only use the max_old_spaces_size argument. I am not >>> sure, but I would >>> either remove the restart argument, or increase it to match >>> max_old_space_size. >>> >>> We also have the "1 rules skipped due to selector errors" constantly, but >>> it doesn't appear to be a limiting factor. I would be happy if it were >>> resolved, however. >>> >>> Perhaps last note, the angular frontend with SSR on consumes a lot of >>> resources. >>> Apparently much more than it previously did, so no matter the settings, >>> you will >>> have to up the CPU and RAM. It could also resolve the proxy timeouts. >>> Our cores (before I tuned the settings) were so overwhelmed, we >>> frequently >>> got 504 ERROR (We have the angular behind reverse proxy.. it did not >>> manage to >>> load within 45 or 50 second limit). >>> >>> I wish you a lot of luck as I am afraid you will need it. >>> >>> Best regards, >>> Majo >>> >>> On Tue, Mar 5, 2024 at 11:30 AM Carolyn Sullivan <lrelli...@gmail.com> >>> wrote: >>> >>>> Hello all, >>>> >>>> As you might have seen if you frequent the Technical Support channel in >>>> the DSpace Slack >>>> <https://dspace-org.slack.com/archives/C3V628QNN/p1709318454084069>, >>>> we've been encountering high CPU usage in DSpace 7.6 leading to decreased >>>> performance (ie. site unavailability) and we've been having a lot of >>>> errors. We're not entirely sure which errors may be the significant ones >>>> relating to our performance challenges, and would welcome any input from >>>> the community to help us improve our site issues. Also, if you're >>>> encountering similar issues, please do let us know--maybe we're all having >>>> the same problems and can solve these collaboratively. Thanks already to >>>> Tim Donoghue and Mark Wood for their suggestions in the DSpace Slack! >>>> I've >>>> aggregated the responses we've already received on these issues here to >>>> enable us to keep track of suggestions. >>>> >>>> So, a summary of our issues: We've set up our server following best >>>> practices in the Performance Tuning documentation >>>> <https://wiki.lyrasis.org/display/DSDOC7x/Performance+Tuning+DSpace>. >>>> We run everything on a single server with 4 CPU and 12 GB of RAM, which >>>> was >>>> the configuration that worked for our previous version of DSpace (6.3). >>>> Initially, we had pm2 configured in cluster mode with max instances and >>>> max_memory_restart: 500M. With this configuration, the node instances kept >>>> restarting ~every minute and seemed to be monopolizing the CPUs, and >>>> starving the other components. Since then, we have since tuned it down to >>>> 3 instances, ie: >>>> >>>> { >>>> "apps": [ >>>> { >>>> "name": "dspace-ui", >>>> "cwd": "/var/dspace-frontend/", >>>> "script": "dist/server/main.js", >>>> "instances": "3", >>>> "exec_mode": "cluster", >>>> "timestamp": "YYYY-MM-DD HH:mm Z", >>>> "out_file": "log/dspace-ui.log", >>>> "error_file": "log/dspace-ui_error.log", >>>> "merge_logs": true, >>>> "env": { >>>> "NODE_ENV": "production", >>>> "NODE_EXTRA_CA_CERTS": "/etc/ssl/certs/rootCA2.crt" >>>> }, >>>> "max_memory_restart": "1500M", >>>> "node_args": "--max_old_space_size=4096" >>>> } >>>> ] >>>> } >>>> >>>> A review of process with top shows very active node instances despite >>>> low traffic (~ 1 request/second): >>>> >>>> Tasks: 289 total, 5 running, 284 sleeping, 0 stopped, 0 zombie >>>> %Cpu(s): 93.9 us, 3.0 sy, 0.0 ni, 3.0 id, 0.0 wa, 0.0 hi, 0.0 si, >>>> 0.0 st >>>> MiB Mem : 11965.2 total, 440.4 free, 10901.1 used, 623.8 >>>> buff/cache >>>> MiB Swap: 0.0 total, 0.0 free, 0.0 used. 613.9 avail >>>> Mem >>>> PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ >>>> COMMAND >>>> 1504803 dspace 20 0 2328180 1.5g 13936 R 100.0 12.5 10:53.08 >>>> node /v+ >>>> 1506783 dspace 20 0 2620092 1.7g 14024 R 93.8 14.8 9:44.49 >>>> node /v+ >>>> 1506913 dspace 20 0 1383380 586472 14180 R 93.8 4.8 4:57.11 >>>> node /v+ >>>> 1508040 dspace 20 0 733380 141452 36952 R 75.0 1.2 0:00.77 >>>> node /v+ >>>> 781 root 20 0 237020 2536 944 S 6.2 0.0 9:41.79 >>>> vmtoolsd >>>> 1 root 20 0 171488 7176 2492 S 0.0 0.1 0:44.04 >>>> systemd >>>> >>>> Our cache settings are set as follows: >>>> # Caching settings >>>> cache: >>>> ... >>>> serverSide: >>>> debug: false >>>> botCache: >>>> max: 1000 >>>> timeToLive: 86400000 # 1 day >>>> allowStale: true >>>> anonymousCache: >>>> max: 1000 >>>> timeToLive: 10000 # 10 seconds >>>> allowStale: true >>>> >>>> *The main question of our systems analyst (Francois Malric): Is this >>>> level of constantly high CPU usage is normal due to node.js? Or is it >>>> likely that our DSpace is displaying poor performance due to underlying >>>> issues?* >>>> >>>> Here are some examples of the errors we've seen: >>>> >>>> (1) From our DSpace Logs: According to this, we don't have that much >>>> traffic (HTTP 0.96 requests/minute), which would likely be higher if bot >>>> traffic was the issue. Nota bene, our pm2 monitor likely has a bug as >>>> it's >>>> showing the units as req/min; should be req/sec. >>>> >>>> lq Process List qqqqqqqqqqqqqqqqqqqqklqq dspace-ui Logs >>>> qqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqk >>>> x[ 1] dspace-ui Mem: 824 MB xx dspace-ui > The response for ' >>>> https://ruor.uottawa.ca/server/api/core/items/d2d3c x >>>> x[ 2] dspace-ui Mem: 316 MB xx dspace-ui > 1 rules skipped due >>>> to selector errors: x >>>> x[ 3] dspace-ui Mem: 777 MB xx dspace-ui > >>>> .custom-file-input:lang(en)~.custom-file-label -> unmatched x >>>> x[ 0] pm2-logrotate Mem: 45 xx dspace-ui > GET >>>> /handle/10393/19705/simple-search?query=&sort_by=score&order=desc x >>>> x xx dspace-ui > *1 rules skipped >>>> due to selector errors: * x >>>> x xx dspace-ui > >>>> *.custom-file-input:lang(en)~.custom-file-label >>>> -> unmatched* x >>>> x xx dspace-ui > Redirecting from >>>> /bitstreams/e524c49e-5fc2-4e74-b69d-0c890238ab3b/dow x >>>> x xx dspace-ui > GET >>>> /bitstreams/e524c49e-5fc2-4e74-b69d-0c890238ab3b/download 302 x >>>> x xx dspace-ui >* ERROR Error: >>>> Cannot set headers after they are sent to the client x* >>>> x xx dspace-ui > at new NodeError >>>> (node:internal/errors:405:5) x >>>> x xx dspace-ui > at >>>> ServerResponse.setHeader (node:_http_outgoing:648:11) x >>>> x xx dspace-ui > at >>>> ServerResponseService.setHeader (/opt/dspace-frontend/dist/ser x >>>> x xx dspace-ui > at Object.next >>>> (/opt/dspace-frontend/dist/server/9366.js:1:4722) x >>>> x xx dspace-ui > at >>>> ConsumerObserver2.next (/opt/dspace-frontend/dist/server/main. x >>>> x xx dspace-ui > at >>>> SafeSubscriber2.Subscriber2._next (/opt/dspace-frontend/dist/s x >>>> x xx dspace-ui > at >>>> SafeSubscriber2.Subscriber2.next (/opt/dspace-frontend/dist/se x >>>> x xx dspace-ui > at >>>> /opt/dspace-frontend/dist/server/main.js:1:4471483 x >>>> x xx dspace-ui > at >>>> OperatorSubscriber2._this._next (/opt/dspace-frontend/dist/ser x >>>> x xx dspace-ui > at >>>> OperatorSubscriber2.Subscriber2.next (/opt/dspace-frontend/dis x >>>> x xx dspace-ui > code: ' >>>> *ERR_HTTP_HEADERS_SENT*' x >>>> x xx dspace-ui > } >>>> x >>>> x xx dspace-ui > 1 rules skipped due >>>> to selector errors: x >>>> x xx dspace-ui > >>>> .custom-file-input:lang(en)~.custom-file-label -> unmatched x >>>> x xx dspace-ui > Warning >>>> [ERR_HTTP_HEADERS_SENT]: *Tried to set headers after they * x >>>> x xx dspace-ui > GET >>>> /items/d2d3cc05-419a-488e-912e-1ff20ab7a654 200 3281.694 ms - - x >>>> >>>> mqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqjmqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqj >>>> lq Custom Metrics qqqqqqqqqqqqqqqqqqklq Metadata >>>> qqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqk >>>> x Heap Size 709.14 MiB xx App Name dspace-ui >>>> x >>>> x Event Loop Latency p95 xx Namespace default >>>> x >>>> x Event Loop Latency 34.32 ms xx Version N/A >>>> x >>>> x Active handles 10 xx Restarts 38 >>>> x >>>> x Active requests 1 xx Uptime 5m >>>> x >>>> x HTTP 0.96 req/min xx Script path >>>> /opt/dspace-frontend/dist/server/main.js x >>>> x HTTP P95 Latency 4009 ms xx Script args N/A >>>> x >>>> x HTTP Mean Latency 868 ms xx Interpreter node >>>> x >>>> >>>> We'd particularly appreciate feedback on the error messages (bolded >>>> above): >>>> >>>> - 1 rules skipped due to selector errors: >>>> - *Suggestion from Mark Wood: this is probably because we're >>>> using Bootstrap 4 >>>> >>>> <https://stackoverflow.com/questions/67581454/receiving-unmatched-pseudo-class-lang-after-updating-angular>* >>>> - .custom-file-input:lang(en)~.custom-file-label -> unmatched >>>> - ERROR Error: Cannot set headers after they are sent to the client >>>> - Warning [ERR_HTTP_HEADERS_SENT]: Tried to set headers after they >>>> - *Suggestion from Mark Wood: * *The Headers Sent errors >>>> seem to be mostly an annoyance, but the constant dumping of stack >>>> traces >>>> is bloating the log.* >>>> >>>> These errors occur constantly :'( >>>> >>>> *Suggestion from Mark Wood on error messages: The most serious is >>>> probably the proxy errors. It appears that PM2 is closing proxy >>>> connections, probably because there are too many. The machine is simply >>>> being asked to do more work than it can handle in the available time. We >>>> see this too, even after doubling our CPU and memory from levels that were >>>> quite adequate for v6. We are about to throw a big increase in resources >>>> at v7 to see if that helps, as it has at other sites.* >>>> >>>> (2) Example of errors from our Apache Error Log (we run Apache to >>>> proxy, as recommended in documentation): >>>> >>>> [Fri Mar 01 14:51:00.740446 2024] [proxy:error] [pid 1494894:tid >>>> 140510257739520] [client 66.XXX.75.XXX:0] AH00898: Error reading from >>>> remote server returned by /handle/10393/19705/simple-search >>>> [Fri Mar 01 14:53:26.192799 2024] [proxy_http:error] [pid 1494894:tid >>>> 140510257739520] (104)Connection reset by peer: [client 66.XXX.75.XXX:0] >>>> AH01102: error reading status line from remote server localhost:4000 >>>> >>>> Some suggestions we've seen in the DSpace Slack already: >>>> Tim Donoghue: >>>> >>>> - Initial increased bot traffic is common, but tends to decrease >>>> over time >>>> - Review major errors in DSpace/Tomcat/Postgres/etc. logs >>>> - Enable more caching in server-side rendering as that uses the >>>> most CPU in Node.js >>>> - Seconded by Mark Wood: Increasing the caching will likely >>>> reduce the CPU demand but memory demand will increase drastically. >>>> - Mark Wood: In general, DSpace 7.x is much more computationally >>>> expensive than the previous versions >>>> >>>> If you've read this far, thank you so much for your time and >>>> consideration. The wider DSpace community seems to be struggling with >>>> these issues, and we would all welcome your observations on these issues >>>> and suggestions for resolving it. >>>> >>>> Best, >>>> Carolyn Sullivan >>>> >>>> -- >>>> All messages to this mailing list should adhere to the Code of Conduct: >>>> https://www.lyrasis.org/about/Pages/Code-of-Conduct.aspx >>>> --- >>>> You received this message because you are subscribed to the Google >>>> Groups "DSpace Technical Support" group. >>>> To unsubscribe from this group and stop receiving emails from it, send >>>> an email to dspace-tech...@googlegroups.com. >>>> To view this discussion on the web visit >>>> https://groups.google.com/d/msgid/dspace-tech/0e817201-5f4a-4fc1-9781-b462cc97134fn%40googlegroups.com >>>> >>>> <https://groups.google.com/d/msgid/dspace-tech/0e817201-5f4a-4fc1-9781-b462cc97134fn%40googlegroups.com?utm_medium=email&utm_source=footer> >>>> . >>>> >>> -- All messages to this mailing list should adhere to the Code of Conduct: https://www.lyrasis.org/about/Pages/Code-of-Conduct.aspx --- You received this message because you are subscribed to the Google Groups "DSpace Technical Support" group. To unsubscribe from this group and stop receiving emails from it, send an email to dspace-tech+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/dspace-tech/19648e8c-e453-4646-9854-008f6f694aa7n%40googlegroups.com.
