[Distributed systems] Basic of Distributed systems

scaling computer systems > vertical scaling . horizontal scaling

 

vertical scaling = scale up = adding more powerful hardware components in a server

horizontal scaling = scale out = adding more servers

 

 

 

why scale out systems ??? 

1. better performance ( = high throughput & parallel processing -> low latency   )

   + 1MRPS is 1 million requests per second

2. fault tolerance (single server often since they run for 24/7) & high availability

3. cost efficiency (using many cheap servers is less expensive than a few high-end servers)

 

scale out and scale up ... mix up!!!

 

interconnecting servers is crucial (... so cabling is important)

 

system components 

1. servers

- computing servers ... specialize in computing multiple cpus with multiple cores

- storage servers ... can store petabytes of data ... tens of HDD hotswaps  (4 unit = 24 hotswap = 24 disks mount)

- GPU server ... to deal with ml workloads so too expensive ... 1u 4x CPU server * 4 = 4u 10x GPU server

2. switches (with 48 ports)

3. server rack

- space efficiency 

- 1 * 32u = 8 * 4u

4. middleboxes

- hw that performs various functions

 

rack scale systems

- a distributed system where its scale is limited to a single rack

- physical scale

- servers can be any type of server

- ToR switches are key component ... interconnects servers in the same rack and different racks

 

multi rack systems ... data center ,  server room

1. leaf spine

2. 3-tier architecture

>> multipath for high availability and load balancing

 

 

 

+ huawei leaf spine

 

 

multi tenant systems

- in single tenant systems, a single tenant monopolizes all the infrastructure, can run multiple applications (google)

- in multi tenant systems, multiple tenants share the infrastructure (aws)