There is a new Redis cluster setup, one team I know in my company is working on, in order to improve the application data caching based out on Redis. The setup is as follows, a Redis cluster with a Redis master and many slaves, say 40-50 (but can grow more when the application is scaled), one Redis instance per one virtual machine. I was told this setup helps the applications deployed in servers on every virtual machines query the data present in the local Redis instance than querying an instance in the network in order to avoid network latency. Periodically, the Redis master is updated only with whatever data are modified or newly created or deleted (data backed by a relational database), say every 5 seconds or so. This will initiate the data sync operation with all the Redis slave instances. The data-consumers (the application deployed on all the virtual machines) of the Redis (slaves) reads updated values to do processing. Is this approach a correct one to the network latency problem faced by the applications in querying from a Redis instance that is within a data center network? Will this setup not create lots of network traffic when Redis master syncing the data with all its slave nodes?
I couldn't find much answers on this from the internet. Your opinions on this are much appreciated.
The relevance of this kind of architecture depends a lot about the workload. Here are the important criteria:
the ratio between the write and read operations. Obviously, the more read operations, the more relevant the architecture. The main benefit IMO, is not necessarily the latency gains, but the scalability, the extra reliability it brings, and the network resource consumption.
the ratio between the cost of a local Redis access against the cost of a remote Redis access. Do not assume that the only cost of a remote Redis access is the network latency. It is not. On my systems, a local Redis access costs about 50 us (in average, very low workload), while a remote access costs 120 us (in average, very low workload). The network latency is about 60 us. Measure the same kind of figures on your own system/network, with your own data.
Here are a few advices:
do not use a single Redis master against many slave instances. It will limit the scalability of the system. If you want to scale, you need to build a hierarchy of slaves. For instance, have the master replicates to 8 slaves. Each slave replicates to 8 other slaves locally running on your 64 application servers. If you need to add more nodes, you can tune the replication factor at the master or slave level, or add one more layer in this tree for extreme scalability. It brings you flexibility.
consider using unix socket between the application and the local slaves, rather than TCP sockets. If it good for both latency and throughput.
Regarding your last questions, you really need to evaluate the average local and remote latencies to decide whether this is worth it. Note that the protocol used by Redis to synchronize master and slaves is close to the normal client server traffic. Every SET commands applied on the master, will be also applied on the slave. The network bandwidth consumption is therefore similar. So in the end, it is really a matter of how many reads and how many writes you expect.
Related
I was reading about Redis and Apache ignite, both of them are in-memory cache and also act as distributed cache. I was wondering what is in-memory cache? Where is the data stored ? In the memory of local system on which an application is being used or in the memory of server where the application is hosted? How does in-memory caching works?
Example:
An application with ignite cache is running on x IP address and I am using the application on y IP address so cache will stored in memory of x IP address system or y IP address system?
Also What does it mean when we say distributed cache?
The in-memory cache can be thought of as a cache that has performance critical information/data of a database that is shared across requests in an application. There is direct access to data/memory rather than through other mechanisms that enables database related operations to operate with high efficiency inturn increasing the throughput, responsiveness of the system.
In general, in the case of distributed cache based on deployment model, the cache memory can be between database and application in a distributed manner. This cache memory can be distributed between the nodes and shall operate based on the distributed hash table and the type of data. The access of data from cache in respective nodes can in turn shall apply the in-memory cache logic to bring in performance optimization.
Here is an example of how it is achieved with Amazon Elastic Cache
As you can see, the Amazaon Elastic Cache solution has a Cache Engine running in each node which implements the caching protocol/algorithm and the Amazon Elastic Cache can support cache sizes from 6 to 67 GB in a particular node. A DNS name is assigned to each Cache Node while it is created and you need to configure the DNS names of the nodes into the client library that is being used. Once your application invokes the Put or Get requests to the cluster, the library shall algorithmically choose a particular node using the hash function that shall spread the data out across the nodes and also help in fetching the same from the nodes.
A distributed cache partitions/shards your data across multiple cluster nodes.This allows utilizing memory and CPU resources of the entire cluster, and load balance requests. A node is a process that can be running on your physical server, virtual machine, or just be a Kubernetes pod. This article might be helpful to understand the basics.
Usually, an application needs to know the IP address of at least one cluster node to open a connection. Once the connection is opened, you would work with the cluster in a way similar to relational databases - just issue your SQL requests, compute tasks and perform other operations.
Also, watch the In-Memory Computing Essentials for Software Engineers recording that covers most of your questions and introduces you to the essential capabilities of Ignite. There is free instructor-led training that is scheduled from time to time on this topic.
My tests with standalone (single-threaded) Redis show that load from a number of parallel clients can drive Redis CPU usage to 100% (in my memory cache use case).
Starting it in cluster mode and sharding the content to multiple masters is a possible approach for speeding it up, if persistence is turned on.
I have a configuration without persistence (turned off RDB and AOF). Would starting multiple masters help performance (still using the same cummulative amount of RAM)?
Redis is single-threaded, so the performance of a standalone instance is limited by processing power of a single CPU core and the network bandwidth of a single machine. However, Redis is very very fast. So normally the bottleneck is network bandwidth, unless you run lots of slow commands/lua scripts.
If you deploy Redis cluster on multiple machines, the performance should be improved no matter whether the persistence is turned on or off. Since you have more CPU cores, and more network bandwidth.
If you deploy Redis cluster on a single machine (each node listen on a unique port), the performance might be improved. It depends... If the bottleneck is network bandwidth, it won't be improved. On the other hand, if the bottleneck is CPU processing power, the performance should be improved. So, in this case, you should do some benchmark with your specific data, specific environment, and specific commands/lua scripts.
Is it safe to use etcd across multiple data centers? As it expose etcd port to public internet.
Do I have to use client certificates in this case or etcd has some sort of authification?
Yes, but there are two big issues you need to tackle:
Security. This all depends on what type of info you are storing in etcd. Using a point to point VPN is probably preferred over exposing the entire cluster to the internet. Client certificates can also be used.
Tuning. etcd relies on replication between machines for two things, aliveness and consensus. Since a successful write must be committed to at majority of the cluster before it returns as successful, your write performance will degrade as the distance between the machines increases. Aliveness is measured with periodic heartbeats between the machines. By default, etcd has a fairly aggressive 50ms heartbeat timeout, which is optimized for bare metal servers running on a local network. Without tuning this timeout value, your cluster will constantly think that members have disappeared and trigger frequent master elections. This gets worse if both of your environments are on cloud providers that have variable networks plus disk writes that traverse the network, a double whammy.
More info on etcd tuning: https://etcd.io/docs/latest/tuning/
Closed. This question is off-topic. It is not currently accepting answers.
Want to improve this question? Update the question so it's on-topic for Stack Overflow.
Closed 9 years ago.
Improve this question
We are deploying a large scale web application that uses only redis as a data store. I notice the the benchmark of our redis master is around 8000 transactions per second on EC2, far less than the stated benchmarks on dedicated hardware.
I understand that there is a performance penalty for running Redis on a virtual machine like EC2, but I would love some pointers from people who have deployed Redis in production environments on EC2 on what EC2 setup you have found most effective for getting more out of redis.
Thanks.
EC2 is probably not the best environment to run Redis on virtualized hardware, but it is a popular one, and there are a number of points to know to get the best from Redis on this platform.
I'm one of the authors of http://redis.io/topics/benchmarks and http://redis.io/topics/latency which cover most of the topics I present below. This is just a summary of the main points.
Virtualization toll
It is not specific to EC2, but Redis is significantly slower when running on a VM (in term of maximum supported throughput). This is due to the fact for basic operations, Redis does not add much overhead to the epoll/read/write system calls required to handle client connections (like memcached, or other efficient key/value stores). System calls are typically more expensive on a VM, and they represent a significant part of Redis activity (especially in benchmarks). In that conditions, a 50% decrease in term of maximum throughput compared to bare metal is not uncommon.
Of course, it also depends on the quality of the hypervisor. For EC2, Xen is used.
Benchmarking in good conditions
Benchmarking can be tricky, especially on a platform like EC2. One point often forgotten is to ensure a proper configuration for both the benchmark client and server. For instance, do not run redis-benchmark on a CPU starved micro-instance (which will likely be throttled down by Amazon) while targeting your Redis server. Both machines are equally important to get a good maximum throughput.
Actually, to evaluate Redis performance, you need to:
run redis-benchmark locally (on the same machine than the server), assuming you have more than one vCPU core.
run redis-benchmark remotely (from a different VM), on a machine whose QoS configuration is equivalent to the server machine
So you can evaluate and compare performance of the machines and the network.
On EC2, you will have the best results with second generation M3 instances (or high-memory, or cluster compute instances) so you can benefit of HVM (hardware virtualization) instead of relying on slower para-virtualization.
The fork issue
This is not specific to EC2, but to Xen: forking a large process can be really slow on Xen (it looks better with kvm). For Redis this is a big problem if you plan to use persistence: both persistence options (RDB or AOF) require the main thread to fork and launch background save or rewrite processes.
In some cases, fork latency can freeze Redis event loop for several seconds. The more memory managed by the Redis instance, the more latency.
On EC2, be sure to use a HVM enabled instance (M3, high-memory, cluster), it will mitigate the issue.
Then, if you have large memory requirements, and your application can tolerate it, consider running several smaller Redis instances on the same machine, and shard your data. It can decrease the latency due to fork operations to an acceptable level.
Persistence configuration
This is a key point to get good performance from Redis (both on VM and bare metal). So please take the time to carefully read http://redis.io/topics/persistence
If you use RDB, keep in mind the memory copy-on-write mechanism will start duplicating pages once the save background process has been forked off. So you need to ensure there is enough memory for Redis itself, plus some margin to cope with the COW. the amount of extra memory depends on your workload. The more you write in the instance, the more extra memory you need.
Please note writing a file may also consume some memory (because of the filesystem cache), so during a Redis background save, you need to account for Redis memory, COW overhead, and size of the dump file.
The machine running the Redis server must never swap. If it does, the result will be catastrophic. Contrary to some other stores, Redis is not virtual memory friendly.
With Linux, be sure to set sensible system parameters: vm.overcommit_memory=1 and vm.swappiness=0 (or a very low value anyway). Do not use old kernel versions: they are quite bad at enforcing a low swappiness (resulting in swapping when a large file is written).
If you use AOF, review the fsync options. It is a tradeoff between raw performance and durability of the write operations. You need to make a choice and define a strategy.
You also need to get familiar with the EC2 storage options. On some VM, you have the choice between ephemeral storage and EBS. On some others, you only have EBS.
Ephemeral storage is generally faster, and you will probably get less issues than with EBS, but you can easily loose your data in case of disk failure or reboot of the host, etc ... You can imagine putting RDB snapshots on ephemeral storage, and then copying the resulting files to EBS directories, as a tradeoff between performance and robustness.
EBS is remote storage: it may eat the standard network bandwidth allocated to the VM, and impact the maximum throughput of Redis. If you plan to use EBS, consider selecting the "EBS-optimized" option to establish a QoS between the standard network and storage links.
Finally, a very common setup for performance demanding instances with EC2 is to deactivate persistence on the master, and only activate it on a slave instance. It is probably less safe for the data, but it may prevent a lot of potential latency issues on the master.
I've to move a Windows based multi-threaded application (which uses global variables as well as an RDBMS for storage) to an NLB (i.e., network load balancer) cluster. The common architectural issues that immediately come to mind are
Global variables (which are both read/ written) will have to be moved to a shared storage. What are the best practices here? Is there anything available in Windows Clustering API to manage such things?
My application uses sockets, and persistent connections is a norm in the field I work. I believe persistent connections cannot be load balanced. Again, what are the architectural recommendations in this regard?
I'll answer the persistent connection part of the question first since it's easier. All good network load-balancing solutions (including Microsoft's NLB service built into Windows Server, but also including load balancing devices like F5 BigIP) have the ability to "stick" individual connections from clients to particular cluster nodes for the duration of the connection. In Microsoft's NLB this is called "Single Affinity", while other load balancers call it "Sticky Sessions". Sometimes there are caveats (for example, Microsoft's NLB will break connections if a new member is added to the cluster, although a single connection is never moved from one host to another).
re: global variables, they are the bane of load-balanced systems. Most designers of load-balanced apps will do a lot of re-architecture to minimize dependence on shared state since it impedes the scalabilty and availability of a load-balanced application. Most of these approaches come down to a two-step strategy: first, move shared state to a highly-available location, and second, change the app to minimize the number of times that shared state must be accessed.
Most clustered apps I've seen will store shared state (even shared, volatile state like global variables) in an RDBMS. This is mostly out of convenience. You can also use an in-memory database for maximum performance. But the simplicity of using an RDBMS for all shared state (transient and durable), plus the use of existing database tools for high-availability, tends to work out for many services. Perf of an RDBMS is of course orders of magnitude slower than global variables in memory, but if shared state is small you'll be reading out of the RDBMS's cache anyways, and if you're making a network hop to read/write the data the difference is relatively less. You can also make a big difference by optimizing your database schema for fast reading/writing, for example by removing unneeded indexes and using NOLOCK for all read queries where exact, up-to-the-millisecond accuracy is not required.
I'm not saying an RDBMS will always be the best solution for shared state, only that improving shared-state access times are usually not the way that load-balanced apps get their performance-- instead, they get performance by removing the need to synchronously access (and, especially, write to) shared state on every request. That's the second thing I noted above: changing your app to reduce dependence on shared state.
For example, for simple "counters" and similar metrics, apps will often queue up their updates and have a single thread in charge of updating shared state asynchronously from the queue.
For more complex cases, apps may swtich from Pessimistic Concurrency (checking that a resource is available beforehand) to Optimistic Concurrency (assuming it's available, and then backing out the work later if you ended up, for example, selling the same item to two different clients!).
Net-net, in load-balanced situations, brute force solutions often don't work as well as thinking creatively about your dependency on shared state and coming up with inventive ways to prevent having to wait for synchronous reading or writing shared state on every request.
I would not bother with using MSCS (Microsoft Cluster Service) in your scenario. MSCS is a failover solution, meaning it's good at keeping a one-server app highly available even if one of the cluster nodes goes down, but you won't get the scalability and simplicity you'll get from a true load-balanced service. I suspect MSCS does have ways to share state (on a shared disk) but they require setting up an MSCS cluster which involves setting up failover, using a shared disk, and other complexity which isn't appropriate for most load-balanced apps. You're better off using a database or a specialized in-memory solution to store your shared state.
Regarding persistent connection look into the port rules, because port rules determine which tcpip port is handled and how.
MSDN:
When a port rule uses multiple-host
load balancing, one of three client
affinity modes is selected. When no
client affinity mode is selected,
Network Load Balancing load-balances
client traffic from one IP address and
different source ports on
multiple-cluster hosts. This maximizes
the granularity of load balancing and
minimizes response time to clients. To
assist in managing client sessions,
the default single-client affinity
mode load-balances all network traffic
from a given client's IP address on a
single-cluster host. The class C
affinity mode further constrains this
to load-balance all client traffic
from a single class C address space.
In an asp.net app what allows session state to be persistent is when the clients affinity parameter setting is enabled; the NLB directs all TCP connections from one client IP address to the same cluster host. This allows session state to be maintained in host memory;
The client affinity parameter makes sure that a connection would always route on the server it was landed initially; thereby maintaining the application state.
Therefore I believe, same would happen for your windows based multi threaded app, if you utilize the affinity parameter.
Network Load Balancing Best practices
Web Farming with the
Network Load Balancing Service
in Windows Server 2003 might help you give an insight
Concurrency (Check out Apache Cassandra, et al)
Speed of light issues (if going cross-country or international you'll want heavy use of transactions)
Backups and deduplication (Companies like FalconStor or EMC can help here in a distributed system. I wouldn't underestimate the need for consulting here)