Thanks for the great proposal, Zhangjian. On point #3, I suspect it should be fairly straightforward to create a small isolated synthetic test to prove (or disprove) the benefits of this approach. By driving a controlled amount of requests per second, you could see latency, memory, CPU, etc. Ideally, it should show meaningful improvements without much degradation in other metrics. Would you be able to spend some time doing that?
Thanks, Sangjin On Tue, May 21, 2024 at 5:13 AM zhangjian <1361320...@qq.com.invalid> wrote: > Hi, xiaoqiao he, thank you for your reply. > > 1.Currently, the server and client protocols within router can be > implemented by extends existing protocols and adding asynchronous > functionality, so it will not affect existing synchronization protocols. > RouterClientNamenodeProtocolServerSideTranslatorPB > RouterClientProtocolTranslatorPB > RouterGetUserMappingsProtocolServerSideTranslatorPB > RouterGetUserMappingsProtocolTranslatorPB > RouterNamenodeProtocolServerSideTranslatorPB > RouterNamenodeProtocolTranslatorPB > RouterRefreshUserMappingsProtocolServerSideTranslatorPB > RouterRefreshUserMappingsProtocolTranslatorPB > > The following issues have implemented asynchronous callbacks for > Rpc.server, but I have not found any other modules to use related functions > Server HADOOP-11552 HADOOP-17046 > In the implementation of asynchronous Rpc.client, this issue is directly > used > Client HADOOP-13226 > Therefore, I believe that asynchronous routers are safe for modifying the > RPC protocol, RPC server, and client > > 2. Forwarding requests to multiple downstream ns, the synchronous router > handler adds requests from multiple downstream ns to the thread pool > (RouterRpcClient.executorService), and then waits for responses from all > downstream ns before returning. Since threads in the thread pool also > process rpc requests synchronously, similar to a handler, the number of > threads in the thread pool directly affects the performance of > invoiceConcurrent, which in turn affects the performance of the handler. > In asynchronous router implementation, the handler calls invoiceConcurrent > to simply convert a request into multiple requests and add them to the asyn > handler thread pool, which can then process the next request in the call > queue; When a connection thread of a downstream ns receives a response, it > will hand it over to the async response for processing. The async response > thread will determine whether it has received all responses from the > downstream ns. If it does, it will continue to process the response. > Otherwise, the async response thread will process the next response. The > asynchronous router uses CompletableFuture.allOf() to implement > asynchronous invoiceConcurrent, and the handler, async handler, async > response, and connection thread still does not need to wait synchronously. > In addition, synchronous routers not only have drawbacks in multi ns > environments, but also in single downstream ns situations, it is often > difficult to decide how many handlers to set for the router, setting it too > much will waste thread resources, and setting it too small will not be able > to give pressure to downstream ns; Asynchronous routers can push requests > to downstream ns without considering how to set handlers. Asynchronous > routers can also better connect to more downstream storage services that > support the HDFS protocol, with better scalability. > > 3.Since I have not yet deployed asynchronous routers to our own cluster, > there is no performance comparison. However, theoretically, I believe that > asynchronous routers will occupy more memory than synchronous routers. > However, I do not believe that it will occupy a lot, especially since we > can control the maximum number of requests entering the router, as > CompletableFuture is stable and widely used; In other aspects, it should be > far superior to synchronous routers, especially in downstream scenarios > with more ns.If anyone is interested, you can also help to make a > performance comparison > > > 2024年5月21日 11:39，Xiaoqiao He <hexiaoq...@apache.org> 写道： > > > > Thanks for this great proposal! > > > > Some questions after reviewing the design doc (sorry didn't review PR > > carefully which is too large.) > > 1. This solution will involve RPC framework update, will it affect other > > modules and how to > > keep other modules off these changes. > > 2. Some RPC requests should be forward concurrently to all downstream NS, > > will it cover > > this case in this solution. > > 3. Considering there is one init-version implementation, did you collect > > some benchmark vs > > the current synchronous model of DFSRouter? > > Thanks again. > > > > Best Regards, > > - He Xiaoqiao > > > > On Tue, May 21, 2024 at 11:21 AM zhangjian <1361320...@qq.com.invalid> > > wrote: > > > >> Thank you for your positive attitude towards this feature. You can debug > >> the UTs provided in PR to better understand the current asynchronous > >> calling function. > >> > >>> 2024年5月21日 02:04，Simbarashe Dzinamarira <simbadz...@apache.org> 写道： > >>> > >>> Excited to see this feature as well. I'll spend more time understanding > >> the > >>> proposal and implementation. > >>> > >>> On Mon, May 20, 2024 at 7:55 AM zhangjian <1361320...@qq.com.invalid> > >> wrote: > >>> > >>>> Hi, Yuanbo liu, thank you for your interest in this feature, I think > >> the > >>>> difficulty of an asynchronous router is not only to implement > >> asynchronous > >>>> functions, but also to consider the readability and reusability of the > >>>> code, so as to facilitate the development of the community. I also > >> planned > >>>> to do the virtual thread you mentioned at the beginning, virtual > Threads > >>>> can achieve asynchronousization elegantly at the code level, but the > >>>> biggest problem is that it is not easy to upgrade the jdk version, no > >>>> matter in the community or in the actual production environment. > >> Therefore, > >>>> I later used CompletableFuture, which is currently supported by jdk 8, > >> to > >>>> achieve asynchronousization. The router is stateless, and the router > rpc > >>>> process is very clear. Therefore, even if CompletableFuture itself is > >> not > >>>> as readable as the virtual thread, if we design it well, we can make > the > >>>> asynchronous process look very clear. > >>>> > >>>> > >>>>> 2024年5月20日 10:56，Yuanbo Liu <liuyuanb...@gmail.com> 写道： > >>>>> > >>>>> Nice to see this feature brought up. I tried to implement this > feature > >> in > >>>>> our internal clusters, and know that it's a very complicated feature, > >> CC > >>>>> hdfs-dev to bring more discussion. > >>>>> By the way, I'm not sure whether virtual thread of higher jdk will > help > >>>> in > >>>>> this case. > >>>>> > >>>>> On Mon, May 20, 2024 at 10:10 AM zhangjian <1361320...@qq.com.invalid > > > >>>>> wrote: > >>>>> > >>>>>> Hello everyone, currently there are some shortcomings in the RPC of > >> HDFS > >>>>>> router： > >>>>>> > >>>>>> Currently the router's handler thread is synchronized, when the > >>>> *handler* thread > >>>>>> adds the call to connection.calls, it needs to wait until the > >>>> *connection* notifies > >>>>>> the call to complete, and then Only after the response is put into > the > >>>>>> response queue can a new call be obtained from the call queue and > >>>>>> processed. Therefore, the concurrency performance of the router is > >>>> limited > >>>>>> by the number of handlers; a simple example is as follows: If the > >>>> number of > >>>>>> handlers is 1 and the maximum number of calls in the connection > thread > >>>> is > >>>>>> 10, then even if the connection thread can send 10 requests to the > >>>>>> downstream ns, since the number of handlers is 1, the router can > only > >>>>>> process one request after another. > >>>>>> > >>>>>> Since the performance of router rpc is mainly limited by the number > of > >>>>>> handlers, the most effective way to improve rpc performance > currently > >>>> is to > >>>>>> increase the number of handlers. Letting the router create a large > >>>> number > >>>>>> of handler threads will also increase the number of thread switches > >> and > >>>>>> cannot maximize the use of machine performance. > >>>>>> > >>>>>> There are usually multiple ns downstream of the router. If the > handler > >>>>>> forwards the request to an ns with poor performance, it will cause > the > >>>>>> handler to wait for a long time. Due to the reduction of available > >>>>>> handlers, the router's ability to handle ns requests with normal > >>>>>> performance will be reduced. From the perspective of the client, the > >>>>>> performance of the downstream ns of the router has deteriorated at > >> this > >>>>>> time. We often find that the call queue of the downstream ns is not > >>>> high, > >>>>>> but the call queue of the router is very high. > >>>>>> > >>>>>> Therefore, although the main function of the router is to federate > and > >>>>>> handle requests from multiple NSs, the current synchronous RPC > >>>> performance > >>>>>> cannot satisfy the scenario where there are many NSs downstream of > the > >>>>>> router. Even if the concurrent performance of the router can be > >>>> improved by > >>>>>> increasing the number of handlers, it is still relatively slow. More > >>>>>> threads will increase the CPU context switching time, and in fact > many > >>>> of > >>>>>> the handler threads are in a blocked state, which is undoubtedly a > >>>> waste of > >>>>>> thread resources. When a request enters the router, there is no > >>>> guarantee > >>>>>> that there will be a running handler at this time. > >>>>>> > >>>>>> > >>>>>> Therefore, I consider asynchronous router rpc. Please view the > issues: > >>>>>> https://issues.apache.org/jira/browse/HDFS-17531 for the complete > >>>>>> solution. > >>>>>> > >>>>>> And you can also view this PR: > >>>> https://github.com/apache/hadoop/pull/6838, > >>>>>> which is just a demo, but it completes the core asynchronous RPC > >>>> function. > >>>>>> If you think asynchronous routing is feasible, we can consider > >> splitting > >>>>>> this PR for easy review in the future. > >>>>>> > >>>>>> The PDF is attached and can also be viewed through issues. > >>>>>> > >>>>>> Welcome everyone to exchange and discuss! > >>>>>> > >>>> > >>>> > >>>> --------------------------------------------------------------------- > >>>> To unsubscribe, e-mail: common-dev-unsubscr...@hadoop.apache.org > >>>> For additional commands, e-mail: common-dev-h...@hadoop.apache.org > >>>> > >>>> > >> > >> > >> --------------------------------------------------------------------- > >> To unsubscribe, e-mail: hdfs-dev-unsubscr...@hadoop.apache.org > >> For additional commands, e-mail: hdfs-dev-h...@hadoop.apache.org > >> > >> > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: common-dev-unsubscr...@hadoop.apache.org > For additional commands, e-mail: common-dev-h...@hadoop.apache.org > >
