This blog post does a nice job in summarizing the difference between three data centers based on various parameters. I have summarized the same as a table:

# Neat Slides on Evolution of Internet

A very good high level introduction to how the Internet evolved. Nice read.

# Marketplaces..

An awesome introduction by a16z.

# On Formal Reasoning

As I am going through Introduction to Algorithms course from MIT, just past the first lecture, we can see the importance of formally proving an algorithm correct. If an algorithm is intuitively incorrect at first, then the need for formal reasoning becomes even more important.

Coming to the 2-d peak finding algorithm, here is a semi-formal reasoning:

- When we find a global maximum in a column and shift to either left (right), it is because this global maximum is less than the left (right) element. So, if there is a new global maximum in this column, then it is definitely more than all the elements in the previous column because it is more than the left (element) itself. So, the algorithm works.
- Will the same argument apply if we find a 1-D peak in a column instead of a global maximum? Does not look to me like so…

# Theta, O and Omega…

The first recitation lecture of Introduction to algorithms gives a different introduction to the asymptotic complexity.

- Theta (n) is when you can get a function g(x) and constants m and n such that our candidate complexity function f(x) is between m.g(x) and n.g(x) – i.e Theta(n) can be said if we can find a function such that it can serve as both tight upper and lower bounds of the candidate function
- When such a function is not possible (lets say we have a sin(x) in the function, per lecture), then we can only find a function g(x) such that f(x) is not more than m.g(x) (there is no “not less than” clause) here. When this is the case, we have O(n)
- When there are both lower bounds and upper bounds for the candidate function f(x), but the lower and upper bounds are not with in a constant factor away, then Theta cannot be used. We can use Omega(lb(x)) or O(ub(x)) – where lb(x) and ub(x) are the lower and the upper bound functions in x respectively.

# Divide and Conquer Approach to Peak Find

I was listening to the peak find lecture from MIT and was confused as to why divide and conquer approach would work as the array is not sorted. It took me some time and a few examples to understand what is going on.

- The comparison at the mid point of array is not same as that of searching, but to check for the peak
- When we branch to the left (or right), we have already checked one side for the property of peakness, the other side remains to be checked.
- if the left (right) element is the only element, then it is the peak (trivial case)
- if not, we further sub-divide the problem and continue on the pieces with the same logic.

Basically, by having “greater than or equal to” in the definition of the peak, we are guaranteed to find a peak in any given array.

# Linux Parameters to Tune for Large Concurrent Connections

MigratoryData has achieved 12 million concurrent connections, that to with a JVM!! Time to rethink about JVM. Some Linux parameters that are tuned are:

- The maximum number of open files the system can handle (by tuning /proc/sys/fs/nr_open)
- The maximum number of open files a process can have (by tuning with “ulimit -n”)
- Increase the number of ephemeral ports on a box (the default being from 32768 to 61000) from “500 to 65535”, with “sysctl -w net.ipv4.ip_local_port_range”. On the server side, it is the same listening address and the listening port that is used to pair with each client end point address.
- Once a connection is established, the quadruple that is used to uniquely identify a connection is <serverIP, serverListeningPort, clientIp, clientEphemeralPort>
- We dont need so many port numbers on the server side, we need only so many open sockets

- Reuse TCP’s TIME_WAIT sockets with “/proc/sys/net/ipv4/tcp_tw_recycle”
- Amount of memory allocated to tcp/udp with “sysctl -w net.ipv4.[tcp|udp]_mem”
- Balance the interrupt processing by changing the smp affinity of each tx/rx queue (now is the time to read more about the Intel 10G ethernet adapter) to a different core (writing the bit to /proc/irq/<interruptNum>/smp_affinity

One thing to note is that with all this connections, CPU was not the bottleneck, but memory was. In fact, CPU was still at around 47% when the server reached 12M connections, but memory reached its brim.

Besides, some JVM tuning is also done, which is out of scope of my mind.🙂. All in all, they broke the C10M benchmark with a stock kernel properly tuned!! Which is a magnificient achievement.