Storage I/O in the era of "many": Many applications, many cores, many devices
Scaling I/O on modern systems that use multicore processors and new storage device technologies such as SSDs is emerging as a key problem in storage I/O.
Stored data is increasing at an alarming rate. Published reports evaluate the trends in data growth and show that by 2020 we will need to cope with 35 Zeta Bytes of stored information [IDC‘2011]. As a result, computing from system memory is no longer feasible, making I/O a main bottleneck for many modern applications.
Interference in the I/O path
Interference at all levels of modern systems across applications, cores, and devices hinders scaling to the required levels of performance with implications on total cost of ownership and eventually missed opportunities:
- Non-critical, high I/O volume applications, such as backup, can dramatically impact the performance of critical applications by 10-100x.
- VM execution costs 10-100x for I/O intensive workloads.
- Sharing devices across applications results in interference at the I/O request level reducing device performance by 10x.
- Understanding and resolving I/O bottlenecks today is becoming a daunting task.